Diffusion models have emerged as a scalable paradigm to learn rich high-dimensional data distributions and generate new samples from them. However, many practical problems require more than generating samples from the learned distribution: inverse problems require samples that explain measurements, visual editing demands control over what changes, large-scale applications demand efficiency. This thesis extends diffusion models to meet these demands, organized around two complementary questions: (1) Can diffusion models provide an effective new route to solving scientific inverse problems, and how? (2) How can diffusion models be made more efficient and controllable for practical generation?
\r\n\r\nFor the first, diffusion models offer what classical priors cannot \u2014 a rich, learned prior over high-dimensional solution spaces; we test whether this translates into an effective route by curating five benchmark problems across distinct scientific domains. The results reveal their promise but also clear gaps, motivating algorithms that handle challenging forward models, numerical stability, and uncertainty. We then develop an inference algorithm that balances between two dynamics, one for the diffusion prior and one for measurement consistency. It produces samples that resolve structure the measurements leave undetermined, while still explaining those measurements with calibrated uncertainty.
\r\n\r\nFor the second, the standard diffusion formulation leaves usable structure untapped; we reshape the model through representation design to exploit it. That structure takes three forms \u2014 the temporal smoothness of the sampling trajectory, the spatial redundancy of image data, and the layered composition of edits. Learning representations of each lets us generate in a single step instead of hundreds, train better diffusion models with far less compute, and enable content-preserving editing.\r\nTogether, this thesis develops diffusion models into scientific inference tools and efficient, controllable generators through algorithm and representation design.
", "doi": "10.7907/ew8b-ey95", "publication_date": "2027", "thesis_type": "phd", "thesis_year": "2027" }, { "id": "thesis:18914", "collection": "thesis", "collection_id": "18914", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:09052026-201047533", "type": "thesis", "title": "From Classical Mechanics to Pattern Recognition: Model-Algorithm Co-Design via Efficient Approximations", "author": [ { "family_name": "Li", "given_name": "Albert Hao", "orcid": "0000-0001-8264-7868", "clpid": "Li-Albert-Hao" } ], "thesis_advisor": [ { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "thesis_committee": [ { "family_name": "Chung", "given_name": "Soon-Jo", "orcid": "0000-0002-6657-3907", "clpid": "Chung-Soon-Jo" }, { "family_name": "Culbertson", "given_name": "Preston", "orcid": "0000-0002-1403-8697", "clpid": "Culbertson-Preston" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Exact descriptions of real systems are rarely computationally tractable, and tractable descriptions are rarely exact. The manner in which one approximates system behavior, computational outcomes, or patterns in data is therefore of primary concern. In turn, the choices made during such approximations dictate which algorithms exist for system analysis, how fast they run, and what guarantees they can carry. The form this design problem takes depends on the origin of the model. When the governing laws are known, an approximation can be constructed deliberately, and the designer chooses which structure it exposes. When only a numerical surrogate is available, its errors are inherited rather than chosen, and the burden shifts to algorithms that can absorb them. When no known governing principles exist at all, the approximation itself must be learned from data, and it is valuable only insofar as it serves the algorithms that consume it. This thesis develops methods exploring all these regimes, beginning from highly-specific challenges motivated by mechanical systems in robotic manipulation and ending with general-purpose models of high-dimensional data like natural videos. Across these settings, recurring themes include computational tractability, robustness to uncertainty, the degree to which physical priors can be leveraged, and the question of how systems and data should be represented. By designing each approximation around the algorithm that consumes it, this thesis achieves state-of-the-art results in a variety of fields ranging from robotic manipulation to video understanding. These include the fastest multi-finger grasp synthesis method of its time, the first real-time sampling-based controllers competitive with reinforcement learning for contact-rich manipulation, and video-text encoders that outperform models trained on significantly more video data.", "doi": "10.7907/eg1k-ss21", "publication_date": "2027", "thesis_type": "phd", "thesis_year": "2027" }, { "id": "thesis:18729", "collection": "thesis", "collection_id": "18729", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06012026-033232974", "primary_object_url": { "basename": "Thesis_Carilli_Maria.pdf", "content": "final", "filesize": 40294216, "license": "cc_by_nc_nd", "mime_type": "application/pdf", "url": "/18729/2/Thesis_Carilli_Maria.pdf", "version": "v6.0.0" }, "type": "thesis", "title": "Genetic Interrogation of Expression Regulation", "author": [ { "family_name": "Carilli", "given_name": "Maria Theresa Natalina", "orcid": "0000-0002-8977-7224", "clpid": "Carilli-Maria-Theresa-Natalina" } ], "thesis_advisor": [ { "family_name": "Pachter", "given_name": "Lior S.", "orcid": "0000-0002-9164-6231", "clpid": "Pachter-L" } ], "thesis_committee": [ { "family_name": "Phillips", "given_name": "Robert B.", "orcid": "0000-0003-3082-2809", "clpid": "Phillips-R" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Engelhardt", "given_name": "Barbara E.", "orcid": "0000-0002-6139-7334", "clpid": "Engelhardt-Barbara-E" }, { "family_name": "Pachter", "given_name": "Lior S.", "orcid": "0000-0002-9164-6231", "clpid": "Pachter-L" } ], "local_group": [ { "literal": "div_bbe" } ], "abstract": "Understanding the effects of variation in the genome on organisms' phenotypes is a central goal of biology. For the past several decades, this has been done by performing large-scale association tests to identify links between genetic variants across the genome and particular diseases or traits. As most variants are in non-coding regions of the genome and act only in specific contexts, there arose the intermediate step of associating variants with gene expression patterns in particular tissues using bulk RNA-sequencing data and, with the recent widespread adoption of single-cell RNA-sequencing, particular cell types. However, associative tests are generally restricted to variants within some sequence distance of the gene, as testing tens of millions of distal variants against tens of thousands of genes in hundreds of contexts is computationally and statistically burdensome. Moreover, associating these proximal variants to changes in average gene expression using scRNA-seq data is far from pinpointing the cellular process they may affect. To move beyond analysis of averages, recent work has advocated for a mechanistic approach to analyzing scRNA-seq by modeling the underlying biological processes of transcription, splicing, and degradation.
\r\n\r\nIn this thesis, we first show how biophysical models are useful for identifying changes in biophysical processes not accessible at the level of mean expression. We next develop accelerated inference procedures for these models to make feasible their application at the scale required for genetic association tests. We then propose a framework for testing for the presence or absence of proximal or distal gene regulation using homozygous crosses. Finally, we couple the genetic testing framework and biophysical models to identify regulatory strategies of biophysical processes across cell types in eight tissues of eight genetically diverse mouse strains.
", "doi": "10.7907/qnzh-rf18", "publication_date": "2026", "thesis_type": "phd", "thesis_year": "2026" }, { "id": "thesis:18750", "collection": "thesis", "collection_id": "18750", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06012026-203154325", "primary_object_url": { "basename": "HL_thesis.pdf", "content": "final", "filesize": 12408723, "license": "other", "mime_type": "application/pdf", "url": "/18750/1/HL_thesis.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Leveraging Structural Uncertainty for Decision Making: from Classical Methods to Foundation Model Agents", "author": [ { "family_name": "Liu", "given_name": "Hao", "orcid": "0009-0005-4006-2585", "clpid": "Liu-Hao" } ], "thesis_advisor": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "thesis_committee": [ { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" }, { "family_name": "Bouman", "given_name": "Katherine L.", "orcid": "0000-0003-0077-4367", "clpid": "Bouman-K-L" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Wang", "given_name": "Yixin", "clpid": "Wang-Yixin" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Uncertainty is fundamental to machine learning and decision-making, but the appropriate way to model it depends on the type of risk or randomness being considered, which in turn is governed by the structure of the specific problem at hand. This thesis explores structural uncertainty across diverse decision-making settings, from classical frameworks such as contextual bandits and linear quadratic control to modern foundation model systems including tool-using LLM agents and automated scientific discovery, arguing that effective decision-making requires identifying and exploiting the structure underlying uncertainty. We begin with off-policy evaluation in contextual bandits, where distributional uncertainty has factorization structure. We introduce a distributionally robust approach that models the factorization of the data-generating process for improved uncertainty quantification and robustness under general covariate shift. Next, we study linear quadratic control with untrusted ML predictions, where system disturbances have compositional structure arising from heterogeneous latent sources. We introduce DISC, an online policy that simultaneously disentangles latent disturbance structure and learns per-component confidence parameters, achieving near-optimal consistency when predictions are accurate while maintaining worst-case robustness guarantees when they fail. For tool-using LLM agents, standard logit-based uncertainty estimation fails to capture agent-environment interaction structure. We develop PROBECAL, which probes LLM hidden states to extract structural features of execution traces, yielding calibrated probabilities over execution correctness and improving existing LLM reasoning techniques. For LLM-based symbolic regression, uncertainty structure lies in the environment feedback mechanism, where scalar metrics like MSE compress rich statistical structure into a single number. We develop PROAUG, enabling code-based data analysis that lets agents actively extract structural signals during evolutionary search, significantly improving discovery efficiency. Across these settings, a common principle emerges: uncertainty is inherently structured, and decision-making systems that explicitly model this structure achieve superior robustness, efficiency, and reliability. This thesis demonstrates that structural uncertainty provides a unifying foundation for trustworthy machine learning across classical and modern AI paradigms.", "doi": "10.7907/mkbv-bt06", "publication_date": "2026", "thesis_type": "phd", "thesis_year": "2026" }, { "id": "thesis:18651", "collection": "thesis", "collection_id": "18651", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:05272026-201026026", "primary_object_url": { "basename": "Qilin_Li_2026.pdf", "content": "final", "filesize": 41858309, "license": "other", "mime_type": "application/pdf", "url": "/18651/1/Qilin_Li_2026.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "From Cells to Functional Tissue Units: Scale-Aware Deep Learning for Biological Microscopy Segmentation", "author": [ { "family_name": "Li", "given_name": "Qilin", "orcid": "000-0002-7191-8965", "clpid": "Li-Qilin" } ], "thesis_advisor": [ { "family_name": "Van Valen", "given_name": "David A.", "orcid": "0000-0001-7534-7621", "clpid": "Van-Valen-D" } ], "thesis_committee": [ { "family_name": "Gkioxari", "given_name": "Georgia", "orcid": "0009-0009-6128-109X", "clpid": "Gkioxari-Georgia" }, { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Van Valen", "given_name": "David A.", "orcid": "0000-0001-7534-7621", "clpid": "Van-Valen-D" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Biological microscopy turns cells, tissues and molecular labels into images, but quantitative biology depends on converting those images into reliable objects. Segmentation defines the cell boundaries, nuclei, functional tissue units (FTUs) and background regions from which morphology, spatial neighborhoods, tissue organization, and disease-associated microenvironments are measured. This step remains difficult because biological objects are dense, small, heterogeneous in appearance, expensive to annotate, and often embedded in gigapixel-scale images with substantial stain and scale variation. General computer-vision models provide useful architectures and representations, but they must be adapted before they can reliably support biological image analysis.
\r\n\r\nThis thesis addresses these challenges through four connected contributions. First, it establishes a cell preprocessing and evaluation workflow including microscopy annotation conversion, tiling, normalization, small-object mask fidelity, and object-level evaluation. Second, it studies Equivariant Mask2Former, showing that finite rotation/reflection-equivariant feature processing can help in limited-label or weak-augmentation settings, but carries substantial computational cost. Third, it contributes to CellSAM, a detector-guided cell segmentation framework in which CellFinder supplies automatic bounding-box prompts for SAM-based mask decoding across diverse cellular imaging modalities. Fourth, it introduces FTURader, a scale-dynamic FTU segmentation pipeline that combines H&E stain normalization, FTUFinder detection, dynamic scale control, and SAM2-based mask decoding.
\r\n\r\nTogether, these contributions bridge cell-scale and FTU-scale microscopy segmentation. They show that reliable biological segmentation depends on controlled data representation, metrics matched to the biological question, prompt generators adapted to dense microscopy, and scale-aware inference when the target changes from cells to anatomy. The resulting perspective is that segmentation should be designed and evaluated as a measurement system for specific biological units, not simply as a mask-prediction task.
", "doi": "10.7907/khhp-yg55", "publication_date": "2026", "thesis_type": "phd", "thesis_year": "2026" }, { "id": "thesis:18375", "collection": "thesis", "collection_id": "18375", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:02182026-000452469", "primary_object_url": { "basename": "yeh_christopher_2026.pdf", "content": "final", "filesize": 8815556, "license": "other", "mime_type": "application/pdf", "url": "/18375/1/yeh_christopher_2026.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Learning Decision-Focused Uncertainty Representations: Theory and Applications in Sustainability", "author": [ { "family_name": "Yeh", "given_name": "Christopher Tzong-Ran", "orcid": "0000-0002-7624-6168", "clpid": "Yeh-Christopher-Tzong-Ran" } ], "thesis_advisor": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" } ], "thesis_committee": [ { "family_name": "Low", "given_name": "Steven H.", "orcid": "0000-0001-6476-3048", "clpid": "Low-S-H" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Donti", "given_name": "Priya L.", "orcid": "0000-0002-8503-7464", "clpid": "Donti-Priya-L" }, { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Managing risks arising from decision-making under uncertainty is a paramount challenge in many sustainability and energy applications. The threat of climate change paired with the increasing complexity of energy systems has made it more important than ever to develop artificial intelligence (AI) and machine learning (ML) methods that can address these challenges. However, standard AI and ML methods suffer from a trade-off between accuracy and reliability. While modern AI models are increasingly capable and accurate, they generally lack rigorous uncertainty quantification for reliability guarantees. On the other hand, traditional methods for uncertainty quantification often rely on strong assumptions and are not designed to optimize decision quality. As a result, existing methods for learning uncertainty representations are often not well-suited for real-world decision-making applications.
\r\n\r\nThis thesis develops new methods for overcoming this trade-off by learning uncertainty representations that are directly useful for decision-making. The first part of the thesis studies batch decision problems, where a model is trained on historical data and then used to make one-shot decisions. A central theme is that good predictions alone are not enough: the uncertainty must be shaped to match the downstream task. Instead, we develop a general framework for decision-focused learning of uncertainty representations, which can be tailored to a wide range of optimization problems and risk guarantees. By leveraging differentiable optimization, we develop end-to-end methods for learning calibrated confidence scores for achieving tail-risk guarantees, uncertainty sets for robust optimization, and generative models for Wasserstein distributionally robust optimization. We also extend decision-focused learning to diffusion models so that full predictive distributions can be used for stochastic optimization. Across energy storage, generator scheduling, and related applications, these methods improve robustness and decision quality while retaining theoretical guarantees or tractable optimization.
\r\n\r\nThe second part of the thesis turns to sequential decision-making. For voltage regulation in distribution grids with uncertain topology, the thesis develops an online control algorithm with a finite-mistake guarantee, showing that safe control can be achieved even while the system is still being learned. The thesis also introduces SustainGym, a benchmark suite of reinforcement learning environments for sustainability applications, including electric vehicle charging, electricity markets, datacenters, cogeneration, and building control. SustainGym exposes realistic distribution shifts, physical constraints, and multi-agent interactions that are often absent from standard benchmarks, and it reveals how brittle off-the-shelf reinforcement learning (RL) methods can be in these settings. We develop a new family of distributionally robust RL algorithms that can learn policies with strong out-of-distribution generalization guarantees, and we show that these methods can achieve superior performance in SustainGym and game-playing environments.
\r\n\r\nTaken together, the thesis argues for a general principle: uncertainty should be learned, calibrated, and evaluated in terms of the decisions it enables. By connecting uncertainty quantification, optimization, online learning, and reinforcement learning, this work provides both theory and practical tools for building reliable decision systems in sustainability-critical applications.
", "doi": "10.7907/h9fw-bq37", "publication_date": "2026", "thesis_type": "phd", "thesis_year": "2026" }, { "id": "thesis:18608", "collection": "thesis", "collection_id": "18608", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:05222026-052420973", "primary_object_url": { "basename": "AFG_Thesis_Final.pdf", "content": "final", "filesize": 20530014, "license": "other", "mime_type": "application/pdf", "url": "/18608/1/AFG_Thesis_Final.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Uncovering Hidden Structure in Data and the Physical World for Seismic Tomography and Beyond", "author": [ { "family_name": "Gao", "given_name": "Angela Fang", "orcid": "0000-0001-8574-8728", "clpid": "Gao-Angela-Fang" } ], "thesis_advisor": [ { "family_name": "Bouman", "given_name": "Katherine L.", "orcid": "0000-0003-0077-4367", "clpid": "Bouman-K-L" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Ross", "given_name": "Zachary E.", "orcid": "0000-0002-6343-8400", "clpid": "Ross-Z-E" }, { "family_name": "Tropp", "given_name": "Joel A.", "orcid": "0000-0003-1024-1791", "clpid": "Tropp-J-A" }, { "family_name": "Bouman", "given_name": "Katherine L.", "orcid": "0000-0003-0077-4367", "clpid": "Bouman-K-L" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Reliable and calibrated imaging strategies are essential for tackling challenging real-world imaging problems. Developing these strategies involves interdisciplinary approaches that translate inverse problem tooling into avenues for real scientific insight. However, raw scientific measurements present many challenges that modern imaging techniques do not adequately address. For instance, many state-of-the-art imaging methods struggle with seismic imaging problems due to several factors including non-linear forward models, uncalibrated noise, sensitivity to initial conditions, and limited angular observations. This thesis aims to overcome these challenges by presenting methods that image under uncertainty, leverage unconventional sources of regularization, and reshape the way we solve imaging inverse problems, ultimately enabling new scientific discoveries.", "doi": "10.7907/vcsd-dn14", "publication_date": "2026", "thesis_type": "phd", "thesis_year": "2026" }, { "id": "thesis:18580", "collection": "thesis", "collection_id": "18580", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:05152026-160124985", "primary_object_url": { "basename": "yu_changhua_2026.pdf", "content": "final", "filesize": 104254454, "license": "other", "mime_type": "application/pdf", "url": "/18580/1/yu_changhua_2026.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Understanding Kinase-Substrate Interaction with Deep Learning and High-Throughput Scanning", "author": [ { "family_name": "Yu", "given_name": "Changhua", "orcid": "0000-0003-4799-4535", "clpid": "Yu-Changhua" } ], "thesis_advisor": [ { "family_name": "Van Valen", "given_name": "David A.", "orcid": "0000-0001-7534-7621", "clpid": "Van-Valen-D" } ], "thesis_committee": [ { "family_name": "Sternberg", "given_name": "Paul W.", "orcid": "0000-0002-7699-0173", "clpid": "Sternberg-P-W" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Mayo", "given_name": "Stephen L.", "orcid": "0000-0002-9785-5018", "clpid": "Mayo-S-L" }, { "family_name": "Van Valen", "given_name": "David A.", "orcid": "0000-0001-7534-7621", "clpid": "Van-Valen-D" } ], "local_group": [ { "literal": "div_bbe" } ], "abstract": "Catalyzed by more than 500 human kinases, protein phosphorylation contributes to almost every aspect of cellular signaling. Despite advances in the past decades, a profound gap persists between the scale of the kinase signaling network and our ability to characterize it: over 95% of human phosphosites lack an assigned kinase, approximately one-third of the human kinome remains functionally understudied, and mechanisms by which kinase domains select substrates remain under-explored. Consequently, the druggable landscape for targeting phosphorylation rewiring events in disease contexts remains limited.
\r\n\r\nThis thesis seeks to address these challenges through developments spanning machine learning, high-throughput interactome screening, and deep mutational scanning. We develop KINBERT, a transformer-based protein language model that jointly encodes paired kinase domain and substrate peptide sequences and demonstrate its utility through disease variant interpretation and identification of host kinase targets during viral infection. We develop PhosphoPCA, a barcoded yeast-based assay that links kinase\u2013substrate phosphorylation to cell growth for high-throughput pooled profiling of kinase specificity and apply it to identify novel substrates for understudied kinases and to engineer validated live-cell kinase biosensors. We leveraged saturated mutagenesis to enable deep mutational scanning of kinase domains with measurable components of variant fitness from protein stability, phosphorylation activity, and substrate selectivity.
\r\n\r\nTogether, the thesis builds an integrated framework for systematically decoding the kinase\u2013substrate interaction space, providing a diverse set of novel technologies for illuminating the dark kinome, interpreting pathogenic phosphoSNVs, uncovering the mutational effect of kinase variants, and enabling generalizable engineering of kinase biosensors.
", "doi": "10.7907/2m7a-nx91", "publication_date": "2026", "thesis_type": "phd", "thesis_year": "2026" }, { "id": "thesis:17802", "collection": "thesis", "collection_id": "17802", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:12152025-232943789", "type": "thesis", "title": "Beyond Text: The Rudiments of Next Generation Foundation Models", "author": [ { "family_name": "Talukder", "given_name": "Sabera", "clpid": "Talukder-Sabera" } ], "thesis_advisor": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Gkioxari", "given_name": "Georgia", "clpid": "Gkioxari-Georgia" } ], "thesis_committee": [ { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Gkioxari", "given_name": "Georgia", "clpid": "Gkioxari-Georgia" }, { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" } ], "local_group": [ { "literal": "div_bbe" } ], "abstract": "This thesis builds non-text and multimodal foundation models that overcome the difficulties of non-text data. These challenges are namely image\u2019s and time series\u2019 (e.g. audio\u2019s and video\u2019s): data heterogeneity, data continuity, and large memory requirements. In order to overcome these attributes we must build models that are information dense, generalizable, and multimodal. By the end of this thesis we will empirically demonstrate that the recipe for performant non-text and multimodal foundation models is: create discrete information dense representations, train models with large scale data in the most generalizable manner possible, and fuse data modalities early in the modeling stack.", "doi": "10.7907/s12m-5692", "publication_date": "2026", "thesis_type": "phd", "thesis_year": "2026" }, { "id": "thesis:17658", "collection": "thesis", "collection_id": "17658", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:08292025-040504302", "primary_object_url": { "basename": "thesis_draft_v3_edited.pdf", "content": "final", "filesize": 46719909, "license": "other", "mime_type": "application/pdf", "url": "/17658/1/thesis_draft_v3_edited.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Advancing Scientific Computational Imaging Through Data-Driven and Physics-Based Priors", "author": [ { "family_name": "Feng", "given_name": "Berthy T.", "orcid": "0000-0002-1843-2165", "clpid": "Feng-Berthy-T" } ], "thesis_advisor": [ { "family_name": "Bouman", "given_name": "Katherine L.", "orcid": "0000-0003-0077-4367", "clpid": "Bouman-K-L" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Bouman", "given_name": "Katherine L.", "orcid": "0000-0003-0077-4367", "clpid": "Bouman-K-L" }, { "family_name": "Freeman", "given_name": "William T.", "orcid": "0000-0002-2231-7995", "clpid": "Freeman-William-T" }, { "family_name": "Gkioxari", "given_name": "Georgia", "clpid": "Gkioxari-Georgia" }, { "family_name": "Daraio", "given_name": "Chiara", "orcid": "0000-0001-5296-4440", "clpid": "Daraio-C" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "The core idea of computational imaging is to supplement limited observable data\r\nwith human-imposed assumptions, or priors. One could formulate a prior as a statistical model or physics model of the object being imaged. However, incorporating such assumptions in the imaging process poses computational challenges, including efficiently expressing sophisticated priors, appropriately balancing priors with observations, and gently enforcing physics constraints. This thesis addresses such challenges with principled methods for bringing informative assumptions into computational imaging. We emphasize applications in scientific imaging, and we focus on two categories of priors as well as the intersection between them: data-driven statistics and physics knowledge.
\r\n\r\nOn the data-driven side, this thesis presents work on score-based priors, including a posterior-estimation method and results of re-imagining the famous M87* black hole from real data with score-based priors. On the physics-based side, we have been able to tackle extremely under-determined imaging problems by enforcing physics constraints, including performing single-viewpoint dynamic tomography of emission near a black hole and characterizing interior material properties from video. As a means towards integrating data-driven and physics-based assumptions, we have developed a method to enforce physics constraints on generative models. In this thesis, we explore each aforementioned project with an emphasis on technical novelty and experimental validation on simulated and real data. By opening new routes for bringing in both data-driven and physics-based assumptions, the methods presented in this thesis enable visualizing scientific phenomena beyond the reach of conventional sensors.
", "doi": "10.7907/cmed-tj81", "publication_date": "2026", "thesis_type": "phd", "thesis_year": "2026" }, { "id": "thesis:17728", "collection": "thesis", "collection_id": "17728", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:10232025-045548517", "type": "thesis", "title": "User-Aligned and Robust Bipedal Locomotion", "author": [ { "family_name": "Li", "given_name": "Kejun", "orcid": "0000-0002-0823-9839", "clpid": "Li-Kejun" } ], "thesis_advisor": [ { "family_name": "Ames", "given_name": "Aaron D.", "orcid": "0000-0003-0848-3177", "clpid": "Ames-A-D" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "thesis_committee": [ { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" }, { "family_name": "Tucker", "given_name": "Maegan L.", "orcid": "0000-0001-7363-6809", "clpid": "Tucker-Maegan-Lindsay" }, { "family_name": "Ames", "given_name": "Aaron D.", "orcid": "0000-0003-0848-3177", "clpid": "Ames-A-D" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_bbe" } ], "abstract": "Bipedal robots are uniquely positioned to operate in environments designed for humans. From humanoids traversing unstructured terrain to robotic exoskeletons assisting individuals with paralysis, these systems highlight the promise of legged locomotion. Yet enabling walking that is both robust and aligned with user needs remains a fundamental challenge, owing to hybrid dynamics, hardware limitations, and the variability across individuals in assistive devices.
\r\n\r\nThe first part of this dissertation addresses user-aligned locomotion. A gait that is theoretically stable may still be rejected in practice if it feels unnatural, uncomfortable, or strenuous. To bridge this gap, we integrate musculoskeletal modeling with trajectory optimization to generate anthropomorphic, dynamically feasible walking gaits, and extend preference-based learning with an active learning formulation that efficiently elicits user feedback within a region of interest while maintaining comfort. Together, these methods enable systematic design of gaits that not only achieve stable walking but also capture the nuanced trade-offs users make between comfort, effort, and naturalness of movement.
\r\n\r\nThe second part of this dissertation focuses on robust locomotion in the face of model mismatch, external disturbances, and environmental variability. We develop robustness strategies spanning multiple layers of the control hierarchy: offline trajectory design informed by robustness metrics grounded in hybrid forward invariance, online adaptation through a data-driven predictive framework, and feedback policies learned in massively parallel simulation using reinforcement learning guided by control Lyapunov functions. While independent, these approaches together provide complementary strategies for handling uncertainty, spanning from offline design to real-time adaptation.
\r\n\r\nAlthough motivated by the challenges of exoskeleton locomotion, the methods are validated on other bipedal platforms such as humanoids and lower-limb prostheses, highlighting their broad applicability to diverse bipedal platforms. Overall, this dissertation shows that principled integration of model-based and data-driven approaches enables locomotion strategies that are robust, adaptive, and aligned with human needs, advancing the deployment of bipedal robots and assistive devices.
", "doi": "10.7907/nfhg-dq67", "publication_date": "2026", "thesis_type": "phd", "thesis_year": "2026" }, { "id": "thesis:17825", "collection": "thesis", "collection_id": "17825", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:01162026-051826165", "primary_object_url": { "basename": "main.pdf", "content": "final", "filesize": 27515998, "license": "other", "mime_type": "application/pdf", "url": "/17825/1/main.pdf", "version": "v3.0.0" }, "type": "thesis", "title": "Constructive Learning for Agile Underactuated Control", "author": [ { "family_name": "Jimenez Rodriguez", "given_name": "Ivan Dario", "orcid": "0000-0002-7230-6733", "clpid": "Jimenez-Rodriguez-Ivan-Dario-" } ], "thesis_advisor": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Ames", "given_name": "Aaron D.", "orcid": "0000-0003-0848-3177", "clpid": "Ames-A-D" } ], "thesis_committee": [ { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Ames", "given_name": "Aaron D.", "orcid": "0000-0003-0848-3177", "clpid": "Ames-A-D" }, { "family_name": "Culbertson", "given_name": "Preston", "orcid": "0000-0002-1403-8697", "clpid": "Culbertson-Preston" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Learning-based methods have achieved remarkable performance in robotic control, yet providing formal guarantees for learned controllers remains an open challenge. The central question motivating this work is how control-theoretic structure can serve as an inductive bias for learning, producing controllers that are both flexible and certifiable.
\r\n\r\nThe key insight is that global properties such as stability and safety reduce to pointwise conditions that neural networks can learn to satisfy. A Lyapunov function certifies stability through a condition that must hold at every state in a region; a barrier function certifies safety through a condition that must hold throughout the safe set. Rather than optimizing over expensive trajectory rollouts, one can sample states and penalize violations of these pointwise conditions. If the conditions are satisfied throughout the relevant region of state space, the global guarantee follows.
\r\n\r\nThree complementary strategies are developed for integrating control structure with learning: training on pointwise certificate conditions, learning inputs to structured controllers that provide guarantees conditional on quantities like uncertainty estimates, and enforcing structure architecturally so that desired properties hold by construction. These strategies are validated on neural ordinary differential equations for certified stability and safety, stereo vision systems that learn perception uncertainty for safe quadruped navigation, and Koopman operator learning with guaranteed linear latent dynamics for predictive control of floating platforms.
\r\n\r\nThe primary application domain is underactuated legged locomotion. Zero dynamics, the residual dynamics that remain when controlled outputs have been driven to zero, provide a framework for coordinating actuated degrees of freedom with passive dynamics. A zero dynamics policy is a learned mapping from unactuated to actuated states that defines a stable invariant manifold. Theoretical results establish existence of such policies for locally controllable systems and prove that manifold stability implies full-state stability. Hardware experiments on a bipedal walker and a hopping robot validate these results: the hopper achieves over three thousand hops across stairs, ramps, and narrow bridges with robust disturbance rejection. These experiments demonstrate that control-theoretic structure and neural network learning can be combined to achieve both formal guarantees and robust physical performance.
", "doi": "10.7907/10aa-vg70", "publication_date": "2026", "thesis_type": "phd", "thesis_year": "2026" }, { "id": "thesis:17815", "collection": "thesis", "collection_id": "17815", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:01132026-015239964", "primary_object_url": { "basename": "Guanzhi_PhD_Thesis.pdf", "content": "final", "filesize": 46786964, "license": "other", "mime_type": "application/pdf", "url": "/17815/1/Guanzhi_PhD_Thesis.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Building Foundation Agents with Internet Knowledge and Large Language Models", "author": [ { "family_name": "Wang", "given_name": "Guanzhi", "orcid": "0009-0004-1921-1796", "clpid": "Wang-Guanzhi" } ], "thesis_advisor": [ { "family_name": "Gkioxari", "given_name": "Georgia", "clpid": "Gkioxari-Georgia" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "thesis_committee": [ { "family_name": "Bouman", "given_name": "Katherine L.", "orcid": "0000-0003-0077-4367", "clpid": "Bouman-K-L" }, { "family_name": "Gkioxari", "given_name": "Georgia", "clpid": "Gkioxari-Georgia" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Fan", "given_name": "Jim", "clpid": "Fan-Jim" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Autonomous agents that perceive, reason, and interact with the world promise a future where intelligent systems assist with daily activities, support work in homes and factories, and take on labor-intensive or repetitive tasks across diverse environments. Tremendous progress has been made across embodied AI, from self-driving cars and autonomous drones to whole-body locomotion and dexterous manipulation, yet building generalist agents that learn continually, generalize broadly, and operate reliably in open-world settings remains an open challenge. Modern systems face three key limitations: the scarcity and cost of collecting large-scale interactive experience; the difficulty of grounding high-level goals into long-horizon behavior; and the fragmentation of perception, language understanding, and motor control across modular, separately trained components.
\r\n\r\nThis dissertation explores a unified and scalable recipe for building foundation agents. The core premise is that internet-scale multimodal data, large language models as the embodied reasoning engine, and unified Vision-Language-Action (VLA) architectures together provide a powerful path toward open-ended embodied intelligence. First, internet-scale multimodal data, including gameplay videos, human activity datasets, online tutorials, and wiki documentation, offers unprecedented breadth, exposing agents to diverse strategies, affordances, and environment configurations that cannot be replicated through controlled robotic data collection alone. Second, large language models provide a flexible cognitive layer for planning, task decomposition, iterative self-improvement, tool use, reward generation, and continual skill acquisition without parameter updates. Third, VLA models unify perception, language grounding, and continuous low-level action, enabling fluid real-time behavior across simulation, gaming environments, and physical robotic platforms, while transferring effectively across task families and embodiments.
\r\n\r\nBy integrating these three components, this dissertation advances the development of embodied agents that can draw upon internet-scale knowledge, reason through language, and translate abstract plans into physical behavior. Together, these ideas aim to move us closer to general, open-ended agents that operate with the reliability, adaptability, and competence required to assist humans across a wide range of real-world tasks.
", "doi": "10.7907/r2ax-s535", "publication_date": "2026", "thesis_type": "phd", "thesis_year": "2026" }, { "id": "thesis:17671", "collection": "thesis", "collection_id": "17671", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:09062025-050711919", "primary_object_url": { "basename": "Thesis.pdf", "content": "final", "filesize": 113472034, "license": "other", "mime_type": "application/pdf", "url": "/17671/2/Thesis.pdf", "version": "v6.0.0" }, "type": "thesis", "title": "Learning to Sample in Computational Imaging: Measurement Acquisition and Posterior Estimation", "author": [ { "family_name": "Wu", "given_name": "Zihui", "orcid": "0000-0002-7622-3548", "clpid": "Wu-Zihui" } ], "thesis_advisor": [ { "family_name": "Bouman", "given_name": "Katherine L.", "orcid": "0000-0003-0077-4367", "clpid": "Bouman-K-L" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Bouman", "given_name": "Katherine L.", "orcid": "0000-0003-0077-4367", "clpid": "Bouman-K-L" }, { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" }, { "family_name": "Song", "given_name": "Yang", "orcid": "0000-0003-3193-1679", "clpid": "Song-Yang" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Many problems in science and engineering require visualizing objects that are not directly observable\u2014such as black holes that are millions of light-years away from the Earth or internal anatomical structures hidden within the human body. Computational imaging is a powerful paradigm that combines sensor design with advanced computational algorithms to make the invisible visible. The typical computational imaging pipeline involves first collecting indirect measurements of the target object and then solving a reconstruction problem. This thesis focuses on two core challenges about sampling along this pipeline: (1) optimizing the sampling process for measurement acquisition, and (2) sampling the posterior distribution of possible reconstructions given noisy measurements.
\r\n\r\nThe first part of the thesis investigates how to design adaptive and task-specific acquisition strategies for computational imaging systems, with a focus on compressed sensing magnetic resonance imaging (CS-MRI). We propose a sequential sampling method that learns to select measurements in multiple stages, and an approach that tailors sampling patterns for specific downstream tasks such as region-of-interest reconstruction, segmentation, and classification. These methods enable a better selection of measurements taken during acquisition, leading to improved performance compared to conventional baselines. We have also implemented our learned sequences on a real MRI scanner and verified their improvement in practice.
\r\n\r\nThe second part of the thesis develops a principled framework for posterior sampling using diffusion models (DMs)\u2014a state-of-the-art class of generative models. By revealing a key connection between DMs and the Split Gibbs Sampling, we introduce a posterior sampling method that rigorously incorporates pre-trained DMs as image priors for solving inverse problems, which exhibits strong performance on a variety of applications. We then show that this framework can be naturally extended into a series of instantiations for solving more general inverse problems, addressing topics like text conditioning, video inverse problems, non-differentiable forward models, and discrete-space sampling. We also present a comprehensive benchmark for systematically evaluating state-of-the-art DM-based posterior estimation methods.
\r\n\r\nBy leveraging machine learning to address challenges in both data acquisition and posterior estimation, this thesis provides new possibilities for building more intelligent and reliable imaging systems across science and engineering.
", "doi": "10.7907/j42z-s192", "publication_date": "2026", "thesis_type": "phd", "thesis_year": "2026" }, { "id": "thesis:18675", "collection": "thesis", "collection_id": "18675", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:05282026-203626843", "primary_object_url": { "basename": "PhD_Thesis.pdf", "content": "final", "filesize": 18975429, "license": "other", "mime_type": "application/pdf", "url": "/18675/1/PhD_Thesis.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Artificial Intelligence Methods for Enzyme Engineering", "author": [ { "family_name": "Yang", "given_name": "Jason", "orcid": "0000-0003-3184-1550", "clpid": "Yang-Jason" } ], "thesis_advisor": [ { "family_name": "Arnold", "given_name": "Frances Hamilton", "orcid": "0000-0002-4027-364X", "clpid": "Arnold-F-H" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "thesis_committee": [ { "family_name": "Wang", "given_name": "Zhen-Gang", "orcid": "0000-0002-3361-6114", "clpid": "Wang-Zhen-Gang" }, { "family_name": "Arnold", "given_name": "Frances Hamilton", "orcid": "0000-0002-4027-364X", "clpid": "Arnold-F-H" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Gradinaru", "given_name": "Viviana", "orcid": "0000-0001-5868-348X", "clpid": "Gradinaru-V" } ], "local_group": [ { "literal": "div_chem" } ], "abstract": "Proteins, such as specialized catalysts called enzymes, offer transformative potential for sustainable chemical synthesis, environmental remediation, and advanced therapeutics. However, engineering proteins for specific industrial or clinical functions remains a formidable challenge due to the expansive and high-dimensional sequence design space. While directed evolution has facilitated significant breakthroughs, the methodology is often constrained by slow iteration cycles and the requirement for a functional starting point. This thesis introduces novel machine learning frameworks, integrated with experimental workflows, to transcend traditional enzyme engineering approaches. We first present Active Learning-Assisted Directed Evolution (ALDE), which employs Bayesian optimization to enable more efficient optimization of protein properties. Afterward, Contrastive Reaction-Enzyme Pretraining (CREEP) is introduced for the annotation and discovery of enzymes with desired \"new-to-nature\" functionalities. Finally, a new paradigm for Steering Generation for Protein Optimization (SGPO) is demonstrated, unifying these two perspectives into a holistic generative framework for efficient protein engineering. Collectively, these innovations advance the transition toward automated, artificial intelligence-driven biomolecular design\u2013unlocking sustainable synthesis, novel therapeutics, and programmable biology at the molecular level.", "doi": "10.7907/wqhh-dm06", "publication_date": "2026", "thesis_type": "phd", "thesis_year": "2026" }, { "id": "thesis:16533", "collection": "thesis", "collection_id": "16533", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:07052024-170119371", "primary_object_url": { "basename": "Thesis.pdf", "content": "final", "filesize": 43525786, "license": "other", "mime_type": "application/pdf", "url": "/16533/1/Thesis.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Active Acquisition Methods for Single Cell Genomics", "author": [ { "family_name": "Chen", "given_name": "Xiaoqiao", "orcid": "0000-0003-4685-3466", "clpid": "Chen-Xiaoqiao" } ], "thesis_advisor": [ { "family_name": "Thomson", "given_name": "Matthew", "orcid": "0000-0003-1021-1234", "clpid": "Thomson-M-W" } ], "thesis_committee": [ { "family_name": "Cai", "given_name": "Long", "orcid": "0000-0002-7154-5361", "clpid": "Cai-Long" }, { "family_name": "Thomson", "given_name": "Matthew", "orcid": "0000-0003-1021-1234", "clpid": "Thomson-M-W" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Bouman", "given_name": "Katherine L.", "orcid": "0000-0003-0077-4367", "clpid": "Bouman-K-L" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "We introduce two novel computational methodologies, ActiveSVM and Active Cell Inference, aimed at reducing the costs and enhancing the efficiency of single-cell mRNA sequencing and spatial transcriptomics, respectively. ActiveSVM employs an active learning approach to identify minimal yet highly informative gene sets for cell-type classification, physiological state identification, and genetic perturbation responses in single-cell datasets. By focusing on misclassified cells through an iterative process, ActiveSVM efficiently scales to analyze over a million cells, demonstrating around 90% accuracy across various datasets, including cell atlas and disease characterization studies.
\r\n\r\nActive Cell Inference complements this by utilizing ordered gene sets, developed through ActiveSVM, to streamline spatial genomics measurements. This end-to-end pipeline significantly reduces measurement time and costs by up to 100-fold in scientific and clinical settings. It optimizes the gene probing process by identifying well-classified cells early, allowing for targeted gene application based on cell classification certainty. This method's efficacy is further enhanced by a temporal scaling calibration scheme, improving calibration accuracy throughout its iterative process.
\r\n\r\nBoth methodologies were rigorously tested on the expansive Human Cell Atlas dataset, using the advanced computational tool, CellxGene-Census, involving over 60 million cells. This integration facilitated the creation of precise gene sets for various human tissues, dramatically improving the efficiency and reliability of these cutting-edge genomic techniques. Together, ActiveSVM and Active Cell Inference represent significant advancements in the application of genomics to clinical diagnostics, therapeutic discovery, and genetic screens, promising substantial reductions in the operational complexities and costs associated with next-generation sequencing technologies.
", "doi": "10.7907/nsn8-nd79", "publication_date": "2025", "thesis_type": "phd", "thesis_year": "2025" }, { "id": "thesis:16653", "collection": "thesis", "collection_id": "16653", "cite_using_url": "https://resolver.caltech.edu/CaltechThesis:08192024-223132153", "primary_object_url": { "basename": "Caltech_Thesis_JY.pdf", "content": "final", "filesize": 7730776, "license": "other", "mime_type": "application/pdf", "url": "/16653/2/Caltech_Thesis_JY.pdf", "version": "v5.0.0" }, "type": "thesis", "title": "Safe and Scalable Learning-Based Control: Theory and Application in Sustainable Energy Systems", "author": [ { "family_name": "Yu", "given_name": "Jing", "orcid": "0000-0003-1318-0189", "clpid": "Yu-Jing" } ], "thesis_advisor": [ { "family_name": "Doyle", "given_name": "John Comstock", "orcid": "0000-0002-1828-2486", "clpid": "Doyle-J-C" }, { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" } ], "thesis_committee": [ { "family_name": "Low", "given_name": "Steven H.", "orcid": "0000-0001-6476-3048", "clpid": "Low-S-H" }, { "family_name": "Doyle", "given_name": "John Comstock", "orcid": "0000-0002-1828-2486", "clpid": "Doyle-J-C" }, { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "Resnick Sustainability Institute" }, { "literal": "div_eng" } ], "abstract": "From intelligent transportation systems to the smart grid, the next generation of cyber-physical systems (CPS) will substantially transform our society. It is vital that these systems are scalable and robust to uncertainties, with contextual awareness and fast adaptation. This dissertation presents progress towards addressing key challenges arising in the control of large-scale CPS, with a special focus on applications in sustainable energy systems.
\r\n\r\nLarge-scale CPS such as the smart grid often consist of numerous interconnected and heterogeneous subsystems that must coordinate to achieve global objectives by exchanging information over a communication network. \r\nTherefore, the first part of this thesis focuses on developing control algorithms that handle crucial design requirements emerging from scalability and communication constraints, such as disturbance localization, communication delay conformation, and distributed implementation.
\r\n\r\nSustainable energy systems are crucial for reducing greenhouse gas emissions and mitigating climate change. However, the inherent unpredictability and large uncertainties associated with renewable generation pose significant challenges for maintaining system stability and safety. Traditional control approaches, while robust and effective for known system models, often fall short when faced with the dynamic and uncertain nature of modern power systems. In the second part of the thesis, we address this challenge by integrating machine learning techniques with model-based control methods using uncertainty sets constructed from real-time data. In particular, we will introduce and provide convergence guarantees for a classic uncertainty set estimation method. Building on these uncertainty sets, we combine learning and control techniques to tackle core CPS control problems, such as adversarial stability certification for linear time-varying systems as well as networked systems under communication constraints where the system models are unknown.
\r\n\r\nThe final part of this thesis applies the developed methodologies to address the voltage control problem in power distribution networks with unknown grid topologies. We will combine online learning techniques and a robust predictive controller to achieve provably finite-time convergence to safe voltage limits, despite uncertainties in network topology and load variations. Our case study on a Southern California Edison 56-bus distribution system demonstrates the effectiveness of this approach in nonlinear, partial observation, and partial control settings.
", "doi": "10.7907/bgar-0602", "publication_date": "2025", "thesis_type": "phd", "thesis_year": "2025" }, { "id": "thesis:16687", "collection": "thesis", "collection_id": "16687", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:09022024-043413501", "primary_object_url": { "basename": "Caltech_Thesis_LaTeX_Template__with_logo_-4.pdf", "content": "final", "filesize": 93184569, "license": "other", "mime_type": "application/pdf", "url": "/16687/1/Caltech_Thesis_LaTeX_Template__with_logo_-4.pdf", "version": "v5.0.0" }, "type": "thesis", "title": "Neurocomputational Understanding of Decision-Making in Novel Environments", "author": [ { "family_name": "Yi", "given_name": "Sanghyun", "orcid": "0000-0003-1274-6523", "clpid": "Yi-Sanghyun" } ], "thesis_advisor": [ { "family_name": "O'Doherty", "given_name": "John P.", "orcid": "0000-0003-0016-3531", "clpid": "O'Doherty-J-P" } ], "thesis_committee": [ { "family_name": "Adolphs", "given_name": "Ralph", "orcid": "0000-0002-8053-9692", "clpid": "Adolphs-R" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Nielsen", "given_name": "Kirby", "orcid": "0000-0003-4536-1021", "clpid": "Nielsen-Kirby" }, { "family_name": "O'Doherty", "given_name": "John P.", "orcid": "0000-0003-0016-3531", "clpid": "O'Doherty-J-P" } ], "local_group": [ { "literal": "div_hss" } ], "abstract": "This thesis investigates the neural and computational mechanisms underlying human decision-making in unfamiliar environments through three interconnected studies. The first study demonstrates that aesthetic value computation for visual art can be systematically predicted from visual features, which are hierarchically represented along the brain's rostrocaudal axis, as revealed by combining deep neural networks with functional MRI data. The second study examines feature-based transfer learning, highlighting the importance of slow integration mechanisms, akin to glial cell functions, for effective knowledge transfer in humans. The third study explores how action affordance influences decision-making in novel environments, showing that action selection results from a competitive interaction between affordance-based and value-based systems, with meta-control exerted by the pre-supplementary motor area and anterior cingulate cortex. Taken together, these studies provide a comprehensive neuro-computational perspective for understanding how the brain navigates novel environments by doing feature-based value computation, transferring knowledge, and using affordance as a guide for action selection.", "doi": "10.7907/akkq-7712", "publication_date": "2025", "thesis_type": "phd", "thesis_year": "2025" }, { "id": "thesis:17295", "collection": "thesis", "collection_id": "17295", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:05292025-064811885", "primary_object_url": { "basename": "PhD_Thesis_Yiheng_Lin.pdf", "content": "final", "filesize": 3464596, "license": "other", "mime_type": "application/pdf", "url": "/17295/3/PhD_Thesis_Yiheng_Lin.pdf", "version": "v5.0.0" }, "type": "thesis", "title": "Predictions and Policy Optimization in Online Decision Making", "author": [ { "family_name": "Lin", "given_name": "Yiheng", "orcid": "0000-0001-6524-2877", "clpid": "Lin-Yiheng" } ], "thesis_advisor": [ { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "thesis_committee": [ { "family_name": "Mazumdar", "given_name": "Eric V.", "orcid": "0000-0002-1815-269X", "clpid": "Mazumdar-Eric" }, { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Srikant", "given_name": "Rayadurgam", "orcid": "0000-0003-1483-5204", "clpid": "Srikant-Rayadurgam" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Predictions are ubiquitous in modern systems, offering insights into how environments might evolve by encoding our prior knowledge and assumptions. Recent advances in artificial intelligence have significantly expanded the scope and accuracy of such models, creating vast new opportunities across domains. At the same time, online decision making remains a fundamental challenge in many real-world problems, concerned with challenges such as limited information, delayed feedback, and irrevocable actions. This dissertation focuses on the interplay between predictions and online decision making---how predictive information can be effectively leveraged to improve performance in dynamic, uncertain environments.
\r\n\r\nWhile incorporating predictions often enhances decision-making, the degree of improvement can vary substantially. This variability arises from two key factors. First, the potential benefit of using predictions is fundamentally determined by both the nature of the predictions (e.g., their targets, errors, and distributions) and the characteristics of the decision-making process (e.g., costs and dynamics). Second, standard predictive policies frequently fall short of realizing such potential, especially in changing environments or when critical system parameters are unknown.
\r\n\r\nThis dissertation introduces a unified theoretical framework to quantify the benefit of leveraging predictions across a broad range of online decision-making problems. To close the gap between the maximum potential and achievable performance, we formulate a general policy optimization framework and design efficient algorithms capable of tracking optimal (predictive) policies in time-varying settings. Additionally, we address practical considerations such as scalability and computational efficiency, enabling the application of our methods in large-scale networks and on resource-constrained devices.
", "doi": "10.7907/37t0-7n77", "publication_date": "2025", "thesis_type": "phd", "thesis_year": "2025" }, { "id": "thesis:17425", "collection": "thesis", "collection_id": "17425", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06092025-020707222", "type": "thesis", "title": "Perception-Driven Autonomy and Learning Control for Ground Vehicles", "author": [ { "family_name": "Lupu", "given_name": "Elena Sorina", "orcid": "0000-0002-3968-2630", "clpid": "Lupu-Elena-Sorina" } ], "thesis_advisor": [ { "family_name": "Chung", "given_name": "Soon-Jo", "orcid": "0000-0002-6657-3907", "clpid": "Chung-Soon-Jo" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Hadaegh", "given_name": "Fred", "orcid": "0000-0002-0992-6323", "clpid": "Fred-Hadaegh" }, { "family_name": "Chung", "given_name": "Soon-Jo", "orcid": "0000-0002-6657-3907", "clpid": "Chung-Soon-Jo" }, { "family_name": "Dabiri", "given_name": "John O.", "orcid": "0000-0002-6722-9008", "clpid": "Dabiri-J-O" }, { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" } ], "local_group": [ { "literal": "GALCIT" }, { "literal": "div_eng" } ], "abstract": "Autonomous robots are widely recognized as highly valuable and are expected to become increasingly prevalent. They will play a critical role across a wide range of terrestrial applications in complex, unstructured environments, as well as in space, supporting infrastructure and exploration on various bodies throughout the solar system and beyond. Looking ahead, autonomous robots will play a crucial role in the search for extraterrestrial life by enabling exploration of remote and extreme environments beyond Earth.\r\nAs robots need to approach more complex tasks, the ability to rapidly perceive, understand, make real-time decisions, and operate at speed requires advances in perception-driven controls, improved predictability, and robustness to disturbances. \r\nTo enable these capabilities, the first part of this thesis proposes an innovative approach to enhancing ground vehicle mobility by integrating a vision-based control algorithm that adapts to changes in real-time. \r\nOur approach improves the vehicle's ability to assess and respond to complex terrains in real-time by leveraging visual information through visual foundation models and meta-learning.\r\nOur controller has provable guarantees of exponential stability and was validated on board two ground vehicles.\r\nNext, an extension of the previously mentioned method applied to detecting objects in space using a visual foundation model is presented. Our method was successfully demonstrated in space in early 2025 aboard the EdgeNode Lite spacecraft.\r\nEfficient operation comes from the synergy of suitable autonomy and control with a suitable robot body.\r\nFollowing this consideration, the second part of the thesis presents the design and control of multi-degrees of freedom robots designed for mobility in complex environments.\r\nIt presents a nonlinear tracking controller with adaptation to improve the walking performance of walking-flying robots. This is illustrated by our implementation on Leonardo, the first robot to combine walking with flying to create a new type of locomotion, which we showcase in complex acrobatic movements such as slacklining and skateboarding.\r\nIn a second case study, we aim to further understand and improve biped walking by introducing a bipedal robot designed to be lightweight, easily manufactured, and easily repaired, serving as a platform for testing learning-based controllers.\r\nWe introduce and demonstrate the performance of two controllers: a model-based and a learning-based control.\r\nThis work highlights the importance of tightly integrated perception, control, and electromechanical design in achieving robust autonomy: on Earth, in orbit, and beyond.", "doi": "10.7907/79tk-eg16", "publication_date": "2025", "thesis_type": "phd", "thesis_year": "2025" }, { "id": "thesis:17406", "collection": "thesis", "collection_id": "17406", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06042025-002343274", "primary_object_url": { "basename": "MeredithHooper_Dissertation.pdf", "content": "final", "filesize": 84549492, "license": "other", "mime_type": "application/pdf", "url": "/17406/1/MeredithHooper_Dissertation.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Machine-Learned Propulsion Strategies: From Adaptive Damage Compensation to Advanced Aeromobility", "author": [ { "family_name": "Hooper", "given_name": "Meredith Leigh", "orcid": "0009-0004-4819-9941", "clpid": "Hooper-Meredith-Leigh" } ], "thesis_advisor": [ { "family_name": "Gharib", "given_name": "Morteza", "orcid": "0000-0003-0754-4193", "clpid": "Gharib-M" } ], "thesis_committee": [ { "family_name": "Dabiri", "given_name": "John O.", "orcid": "0000-0002-6722-9008", "clpid": "Dabiri-J-O" }, { "family_name": "Ames", "given_name": "Aaron D.", "orcid": "0000-0003-0848-3177", "clpid": "Ames-A-D" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Gharib", "given_name": "Morteza", "orcid": "0000-0003-0754-4193", "clpid": "Gharib-M" } ], "local_group": [ { "literal": "GALCIT" }, { "literal": "div_eng" } ], "abstract": "Autonomous vehicles are regularly sent into \"dull, dirty, and dangerous\" environments where the risk of damage is high. Avoidance or mitigation of such damage is therefore paramount to maintain effective autonomy. In this thesis, we use machine learning to investigate two different propulsive strategies that may be used by autonomous vehicles. The first, flapping propulsion, shows remarkable ability in nature to recover from damage simply by altering stroke kinematics. Using machine learning, we ask whether and how such mitigation of damage would be possible for a robotic autonomous vehicle. The second propulsive strategy we investigate is single-rotor propulsion, most commonly seen in helicopters. With this system, we seek to avoid damage before it occurs by improving mobility and control authority via thrust vectoring.
\r\n\r\nIn Part I, we use an evolutionary strategy (CMA-ES) with hardware-in-the-loop to explore optimal machine-learned adaptations to propulsor damage. Experimental function evaluations are performed by a flexible propulsor actuated by a spherical parallel manipulator (SPM). The machine-learned forces and trajectory parameters are compared to in vivo observations in order to determine whether bio-inspired strategies to adapt to significant propulsor damage are the most efficient, or whether they may be affected by irrelevant evolutionary pressures. With amputation of approximately 50% of the propulsor, we find that a complete recovery in thrust production and fitness is made. Some characteristics of the recovered trajectory are similar to natural swimmers, while others differ. Recovery when producing side-force is even more complex. Not all trials are able to recover force production and fitness, and no clear strategy to modify amplitude or frequency is seen. We conclude Part I by using PIV measurements to detail the effect of compensatory strategies on hydrodynamics. Both amputated and intact trajectories clearly show utilization of a drag-based paddling strategy, but the hydrodynamics of the intact and amputated fins differ significantly. This suggests that the machine-learned trajectories are not simply reestablishing the same wake as the intact fin to achieve the same thrust and fitness.
\r\n\r\nGiven the success in applying machine learning in-the-loop to a complex propulsive system where fluid-structure interactions are significant, we utilize the same strategy in Part II to begin to explore helicopter aeromechanics. We built an independent blade control (IBC) system that interfaces with the CMA-ES algorithm to explore optimal blade pitch trajectories. Using this platform, we explore two preliminary optimizations designed to vector thrust; the first, for sustained thrust vectoring that might be utilized upon takeoff or landing, and the second, for short-time thrust vectoring that could be used for enhanced maneuverability. We present some preliminary results from these optimizations and lay out a foundation for future applications of this experimental system.
", "doi": "10.7907/79aa-ja50", "publication_date": "2025", "thesis_type": "phd", "thesis_year": "2025" }, { "id": "thesis:17394", "collection": "thesis", "collection_id": "17394", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06032025-023615578", "primary_object_url": { "basename": "Noel_Csomay_Shanklin_Thesis.pdf", "content": "final", "filesize": 134391515, "license": "other", "mime_type": "application/pdf", "url": "/17394/1/Noel_Csomay_Shanklin_Thesis.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Layered Control Architectures: Constructive Theory and Application to Legged Robots", "author": [ { "family_name": "Csomay-Shanklin", "given_name": "Noel V.", "orcid": "0000-0002-2361-1694", "clpid": "Csomay-Shanklin-Noel-V" } ], "thesis_advisor": [ { "family_name": "Ames", "given_name": "Aaron D.", "orcid": "0000-0003-0848-3177", "clpid": "Ames-A-D" } ], "thesis_committee": [ { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" }, { "family_name": "Ames", "given_name": "Aaron D.", "orcid": "0000-0003-0848-3177", "clpid": "Ames-A-D" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Niemeyer", "given_name": "Gunter", "clpid": "Niemeyer-Gunter" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Fueled in part by the imagination of science fiction, every decade since the 1950s has expected robots to enter our everyday lives in the subsequent decade. Despite this anticipation, the widespread adoption of robots has consistently fallen short of societal expectations. This delay is attributable to the sheer variety of complexities in robotics --- perception, contact-rich dynamics, human-robot interactions. Each sub-discipline of robotics poses unique challenges that must be addressed to achieve general autonomy. As progress is made in these sub-fields, it is increasingly important to adopt a layered architecture perspective that combines isolated controller blocks into a unified framework.
\r\n\r\nThis thesis argues that on the road to general autonomy, adopting layered architectures enables three key benefits: efficiency, feasibility, and generalizability. We root our discussion in a general problem in robotics: the design of a controller that navigates a robot to a goal state while satisfying all state and input constraints that are present. Throughout the thesis, we focus on solutions that are both general --- applicable across a wide variety of robotic platforms --- and concrete --- deployed and tested on specific hardware platforms. As such, we aim to not only propose a framework for reasoning about this problem, but also methods to synthesize controllers that solve it in practice for legged robots.
\r\n\r\nWe begin by motivating and formalizing the notion of layered architectures and use this to build our control stack from the bottom up. We start with low-level planning and tracking layers that stabilize the system within a tracking tube for both the actuated and underactuated states of legged robots. We then introduce high-level planning and tracking layers that generate and follow sparse, dynamically feasible graphs for coarse global navigation through cluttered environments. By decomposing the global control problem into interacting levels and layers, each operating with disparate timescales and system abstractions, we enable tractable, reliable, and extensible robot autonomy.
\r\n\r\nThroughout this thesis, an emphasis will be placed on mathematical structure, constructive synthesis, and experimental validation. We demonstrate that adopting a layered architecture perspective is not merely an implementation convenience, but a fundamental organizing principle that can enable true robot autonomy.
", "doi": "10.7907/k0ns-c606", "publication_date": "2025", "thesis_type": "phd", "thesis_year": "2025" }, { "id": "thesis:17388", "collection": "thesis", "collection_id": "17388", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06032025-000834048", "type": "thesis", "title": "Aligning and Comparing Vision Representations to Improve Understanding and Performance", "author": [ { "family_name": "Kondapaneni", "given_name": "Neehar", "orcid": "0000-0002-2782-3977", "clpid": "Kondapaneni-Neehar" } ], "thesis_advisor": [ { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Gkioxari", "given_name": "Georgia", "clpid": "Gkioxari-Georgia" }, { "family_name": "Mac Aodha", "given_name": "Oisin", "orcid": "0000-0002-5787-5073", "clpid": "Mac-Aodha-Oidin" }, { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" } ], "local_group": [ { "literal": "div_bbe" } ], "abstract": "Recent advances in large artificial intelligence (AI) models have enabled these models to perform a wide range of real-world tasks with skill levels comparable to or surpassing those of humans. In this thesis, we develop methods to compare, analyze, and align data representations from these powerful models. In Part 1, we develop methods for estimating human knowledge during a learning task and for comparing various data representations. These methods are steps towards a system designed to help us learn from AI.
\r\n\r\nIn Part 2, we show how aligning models can be useful in two separate domains. First, we discover and fix a misalignment in the inputs to a powerful foundation model and show how it improves performance. Second, we show that biologically inspired object manipulation tasks can be used as a training signal for learning human-aligned representations of number. Our results demonstrate the potential for alignment and comparison methods to improve the overall performance of AI models, improve our understanding of biological intelligence, and help us discover new patterns in the natural world.
", "doi": "10.7907/0crh-zb71", "publication_date": "2025", "thesis_type": "phd", "thesis_year": "2025" }, { "id": "thesis:17258", "collection": "thesis", "collection_id": "17258", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:05212025-232450900", "primary_object_url": { "basename": "Thesis_Vafidis.pdf", "content": "final", "filesize": 20191559, "license": "other", "mime_type": "application/pdf", "url": "/17258/1/Thesis_Vafidis.pdf", "version": "v8.0.0" }, "type": "thesis", "title": "Neural Network Models of Learning and Generalization", "author": [ { "family_name": "Vafeidis", "given_name": "Panteleimon", "orcid": "0000-0002-9768-0609", "clpid": "Vafeidis-Panteleimon" } ], "thesis_advisor": [ { "family_name": "Rangel", "given_name": "Antonio", "clpid": "Rangel-A" } ], "thesis_committee": [ { "family_name": "O'Doherty", "given_name": "John P.", "orcid": "0000-0003-0016-3531", "clpid": "O'Doherty-J-P" }, { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Rangel", "given_name": "Antonio", "clpid": "Rangel-A" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Neural networks have emerged as powerful models for understanding both biological and artificial intelligence. This thesis investigates fundamental principles of learning and generalization across four interconnected domains, bridging insights from theoretical neuroscience and machine learning to advance our understanding of intelligent systems.
\r\n \r\nChapter I addresses a central question in associative learning: how do neural circuits learn to associate concepts with one another? We combine two cortical inductive biases, namely mixed selectivity and predictive learning in compartmentalized neurons, to explain how the cortical architecture may confer significant evolutionary advantages for efficient learning and packing multiple associations within the same neuronal population. Our model achieves stimulus substitution, where neurons respond identically to a conditioned stimulus as they would to the associated unconditioned stimulus, a feat in which traditional, Hebb-based learning rules fail.
\r\n\r\nChapter II pivots from the static mappings between concepts learned in Chapter I to explore how neural systems develop the precise synaptic connectivity required to establish dynamic mappings for path integration\u2014the ability to maintain an internal sense of location without external cues. Applied to the Drosophila head direction system, our model develops connectivity patterns strikingly similar to those observed experimentally, with Continuous Attractor (CAN) dynamics emerging naturally from learning. This offers a novel perspective on how precisely calibrated neural circuits can develop through experience, rather than requiring genetic pre-specification, and explains experimental findings where animals adapt their internal representation when sensory experience changes.
\r\n\r\nIn Chapter III, we establish a theoretical framework explaining how disentangled representations\u2014internal models that isolate independent factors of variation in the world\u2014emerge from multi-task learning. We prove that any system competent at multiple related tasks must implicitly represent the underlying latent variables in a linearly decodable form. We experimentally confirm all major theoretical predictions, and reveal a fundamental connection between task diversity and representation quality, particularly explaining why modern transformer models may develop human-interpretable concepts. Furthermore, our work suggests that the massively parallel cortical architecture may be a key facilitator in the development of representations that enable the impressive zero-shot generalization ability that humans possess.
\r\n\r\nFinally, Chapter IV proposes leveraging Large Language Models (LLMs) as cognitive tools for evaluating latent factor hypotheses for psychology, leveraging the theoretical insights from Chapter III. It suggests that the self-consistency of an LLM's responses given hypothesized psychological factors could serve as a metric for psychological latent factor hypothesis evaluation. While preliminary, this approach represents a novel computational methodology for psychology that could transform how hypotheses for human cognition are developed and refined.
\r\n\r\nContinuous attractors display prominently in this thesis (Chapters II and III), yet these concepts are misunderstood, particularly in the experimental literature. Hence, in Appendix D of this thesis we provide important considerations about the detection and quantification of Continuous Attractors (CANs) from experimental data, considerations particularly important in order to avoid confusion when it comes to these concepts, leading to wasted efforts and resources in the experimental neuroscience community.
\r\n\r\nTogether, these investigations reveal complementary aspects of how intelligent systems develop useful, generalizable representations through learning. From biologically plausible learning rules to abstract computational principles, this thesis demonstrates how neural networks can illuminate fundamental mechanisms of intelligence across natural and artificial systems, contributing to a unified science of Neural Computation.
", "doi": "10.7907/cxs1-ss10", "publication_date": "2025", "thesis_type": "phd", "thesis_year": "2025" }, { "id": "thesis:17351", "collection": "thesis", "collection_id": "17351", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06022025-015834886", "primary_object_url": { "basename": "RKC_Thesis-7-compressed.pdf", "content": "final", "filesize": 19037198, "license": "other", "mime_type": "application/pdf", "url": "/17351/1/RKC_Thesis-7-compressed.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Dynamic Safety Under Uncertainty: A Control Barrier Function Approach", "author": [ { "family_name": "Cosner", "given_name": "Ryan Kazuo", "orcid": "0000-0002-4035-1425", "clpid": "Cosner-Ryan-Kazuo" } ], "thesis_advisor": [ { "family_name": "Ames", "given_name": "Aaron D.", "orcid": "0000-0003-0848-3177", "clpid": "Ames-A-D" } ], "thesis_committee": [ { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Niemeyer", "given_name": "Gunter", "clpid": "Niemeyer-Gunter" }, { "family_name": "Culbertson", "given_name": "Preston", "orcid": "0000-0002-1403-8697", "clpid": "Culbterson-Preston" }, { "family_name": "Ames", "given_name": "Aaron D.", "orcid": "0000-0003-0848-3177", "clpid": "Ames-A-D" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Modern technological achievements in robotics, machine learning, and control promise an exciting future where autonomous robots are a useful part of everyday life, from automated manufacturing and driverless cars to robotic healthcare and autonomous delivery drones. However, as robots are deployed in increasingly complex, uncertain, and human-interactive environments, safety becomes paramount; we cannot deploy these systems at scale unless we are rigorously assured of their safety. Despite the capabilities of modern robotics, practical real-world safety is often achieved through conservative hardware designs, confining deployment regulations, or restrictive assumptions that severely limit a robot's capabilities.
\r\n\r\nThe goal of this thesis is to develop methods for achieving dynamic safety: formal safety guarantees that preserve system performance and remain valid under uncertainty. To this end, this thesis advances the theory and practice of control barrier functions (CBFs), a leading framework for enforcing safety constraints on dynamical systems. While CBF-based methods offer strong theoretical guarantees, they do so by relying on several restrictive assumptions. Namely, they assume that the safety requirement and the system dynamics are compatible and that the dynamics model and state are perfectly known. These assumptions rarely hold in real-world settings and can result in false confidence and catastrophic safety failures when violated. This thesis addresses these gaps by systematically relaxing these assumptions and developing new theory to retain rigorous, deployable guarantees.
\r\n\r\nBy leveraging structural properties of several relevant classes of system dynamics, I first present a myriad of constructive synthesis methods that make CBF design feasible for a wide range of robots. I then develop robust control methods that retain their safety guarantees in the presence of bounded dynamics and measurement uncertainty. However, despite the utility of these methods in guaranteeing safety, they often lead to highly conservative behavior that compromises system performance. Thus, to mitigate this conservatism, I leverage machine learning techniques to reduce uncertainty and determine desired levels of robustness. While this unification of machine learning techniques with safety-critical control may sacrifice formal guarantees, it enables safe and performant behavior in practice. Moreover, the robust CBF framework provides a valuable degree of interpretability absent from typical end-to-end approaches.
\r\n\r\nNext, seeking a middle ground between conservative absolute guarantees and capable-but-heuristic methods, I adopt a probabilistic notion of safety that provides risk-based guarantees in the presence of unbounded disturbances. In particular, by illustrating the fundamental connection between DCBFs and supermartingales, I develop new theoretical guarantees and propose several algorithms to achieve safety in the presence of stochastic uncertainty. I then deploy these methods on several complex systems experiencing significant uncertainty, including a quadrotor robot with a slung payload, a humanoid robot walking in unstructured environments, and multiple robots performing dynamic collision avoidance. To achieve this, we use generative modeling techniques to capture the necessary understanding of the uncertainty distribution. Here, I also forego the traditional CBF-based safety filter paradigm and show the performance and safety improvements that can be gained through the unification of CBFs and horizon-based methods such as model predictive control (MPC).
\r\n\r\nTogether, the contributions of this thesis represent an advancement towards dynamic, safe, and capable robotic autonomy under uncertainty. The risk-aware, robust safety-critical control methods proposed here help close the gap between theoretical safety guarantees and the demands of real-world deployment.
", "doi": "10.7907/eee7-0m74", "publication_date": "2025", "thesis_type": "phd", "thesis_year": "2025" }, { "id": "thesis:17182", "collection": "thesis", "collection_id": "17182", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:04292025-041809434", "primary_object_url": { "basename": "fzl_thesis_final.pdf", "content": "final", "filesize": 29018197, "license": "other", "mime_type": "application/pdf", "url": "/17182/1/fzl_thesis_final.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Evaluation of the Generalizability of Machine Learning-Assisted Protein Engineering Methods", "author": [ { "family_name": "Li", "given_name": "Francesca-Zhoufan", "orcid": "0000-0002-5710-9512", "clpid": "Li-Francesca-Zhoufan" } ], "thesis_advisor": [ { "family_name": "Arnold", "given_name": "Frances Hamilton", "orcid": "0000-0002-4027-364X", "clpid": "Arnold-F-H" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "thesis_committee": [ { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" }, { "family_name": "Mayo", "given_name": "Stephen L.", "orcid": "0000-0002-9785-5018", "clpid": "Mayo-S-L" }, { "family_name": "Yang", "given_name": "Kevin K.", "orcid": "0000-0001-9045-6826", "clpid": "Yang-Kevin-K" }, { "family_name": "Arnold", "given_name": "Frances Hamilton", "orcid": "0000-0002-4027-364X", "clpid": "Arnold-F-H" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_bbe" } ], "abstract": "Engineered proteins can carry out a vast array of functions and have become indispensable across numerous industrial applications. To accelerate wet-lab protein engineering efforts, machine learning-based methods have advanced rapidly. However, a gap remains between state-of-the-art machine learning methods and their practical adoption. A key factor contributing to this disconnect is the lack of application-relevant benchmarking and generalizable insights across protein engineering tasks. This thesis evaluates machine learning-assisted protein engineering approaches to identify generalizable strategies. The central problem considered is learning the mapping from protein sequence to function\u2014known as the fitness landscape\u2014to enable the prediction of unseen variant fitness. Chapter 1 introduces the background and context for machine learning-assisted protein engineering and highlights the practical constraint of limited experimental budgets. Chapter 2 investigates transfer learning, which leverages models pretrained on large protein sequence databases to generate informative representations for modeling task specific sequence-function relationships. Evaluation across ten diverse tasks shows that while transfer learning is effective in structure prediction, it underperforms in variant fitness prediction\u2014a key objective in protein engineering. Chapter 3 evaluates alternative strategies with a focus on combinatorial fitness landscapes, a common setting in protein engineering. Across 16 diverse landscapes, focused training improves the performance of various machine learning approaches by strategically selecting training variants using zero-shot predictors, which estimate variant fitness from auxiliary information without relying on experimental data. Building on these insights, Chapter 4 addresses the specific challenge of engineering enzymes\u2014proteins that convert substrates into products\u2014for novel chemistries. While six general zero-shot predictors without substrate information can predict enzyme activity on non-native substrates, they fail on more out-of-distribution, new-to-nature chemistries. Incorporating substrate information into zero-shot predictors leads to more generalizable performance across all tested chemistries, spanning 22 substrates. Overall, this thesis identifies generalizable strategies for machine learning-assisted protein engineering by systematically evaluating and improving how sequence-to-function relationships are modeled across diverse tasks.", "doi": "10.7907/yzb2-cb66", "publication_date": "2025", "thesis_type": "phd", "thesis_year": "2025" }, { "id": "thesis:16185", "collection": "thesis", "collection_id": "16185", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:09212023-182422125", "type": "thesis", "title": "Acquiring Enzyme Sequence-Fitness Data at Scale Toward Predictive Methods for Enzyme Engineering", "author": [ { "family_name": "Johnston", "given_name": "Kadina Elizabeth", "orcid": "0000-0002-2214-3534", "clpid": "Johnston-Kadina-Elizabeth" } ], "thesis_advisor": [ { "family_name": "Arnold", "given_name": "Frances Hamilton", "orcid": "0000-0002-4027-364X", "clpid": "Arnold-F-H" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Mayo", "given_name": "Stephen L.", "orcid": "0000-0002-9785-5018", "clpid": "Mayo-S-L" }, { "family_name": "Bois", "given_name": "Justin S.", "orcid": "0000-0001-7137-8746", "clpid": "Bois-J-S" }, { "family_name": "Arnold", "given_name": "Frances Hamilton", "orcid": "0000-0002-4027-364X", "clpid": "Arnold-F-H" } ], "local_group": [ { "literal": "div_bbe" } ], "abstract": "The emergence of machine learning methods for expediting directed evolution via protein fitness prediction has recently shed light on the need for more, high quality sequence-fitness data from which to learn the mapping from sequence to fitness. Enzymes specifically are highly selective catalysts and engineered enzymes are becoming increasingly important for human applications such as pharmaceutical synthesis. This thesis thus focuses on the collection of enzymatic sequence-fitness data to enable both development and validation of emerging approaches. Chapter 1 describes the process of traditional directed evolution as well as ways that machine learning methods have been used to accelerate it. It also discusses the experimental considerations for applying machine learning to the various steps of protein engineering campaigns, as the experimental constraints are not always obvious to the machine learning community. One of the major constraints for the application of machine learning methods is the requirement to sequence all variants required for model training, a step that is often skipped by traditional, lab-only directed evolution due to it not being worth the time and cost. Chapter 2 introduces a solution to this problem with \u201cevery variant sequencing\u201d (evSeq), which enables higher throughput collection of sequencing data for a similar time and cost as commonly used Sanger sequencing methods. This method not only enables implementation of ML methods such as machine learning-assisted directed evolution (MLDE) and focused training MLDE (ftMLDE) by sequencing variants during an evolution campaign, but also offers promise to fill existing protein sequence-fitness databases with protein engineering datasets. This type of data collection can enable the development of newer, more accurate ML methods, and was an inspiration for the work presented in Chapter 3, which details the collection of a combinatorially complete, epistatic sequence-fitness landscape in an enzyme active site. Oftentimes, the effects of mutations on protein fitness can be considered largely independent and laboratory recombination of them can find an optimal variant. This general principle breaks down when the effects of mutations are not independent, termed epistasis, and sequence-fitness landscapes with these interactions are difficult to traverse. Thus, collection of this dataset provides a challenging task for the development of both ML and physics-based models and pushes the boundary of predictive methods for protein engineering.", "doi": "10.7907/xjz4-k217", "publication_date": "2024", "thesis_type": "phd", "thesis_year": "2024" }, { "id": "thesis:16508", "collection": "thesis", "collection_id": "16508", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06062024-061120491", "primary_object_url": { "basename": "Cameron_Voloshin_2024_Thesis.pdf", "content": "final", "filesize": 18301936, "license": "other", "mime_type": "application/pdf", "url": "/16508/1/Cameron_Voloshin_2024_Thesis.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Guaranteed Policy Performance in Reinforcement Learning", "author": [ { "family_name": "Voloshin", "given_name": "Cameron", "orcid": "0009-0007-7725-6660", "clpid": "Voloshin-Cameron" } ], "thesis_advisor": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "thesis_committee": [ { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Bouman", "given_name": "Katherine L.", "orcid": "0000-0003-0077-4367", "clpid": "Bouman-K-L" }, { "family_name": "Chaudhuri", "given_name": "Swarat", "orcid": "0000-0002-6859-1391", "clpid": "Chaudhuri-Swarat" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Decision-making is ubiquitous in everyday life. Increasingly, researchers are seeking answers on how to optimally solve sequential decision-making tasks. Thanks to recent availability of computation, advances in deep learning, and released open-sourced code, it has become easy to train a computational agent to make decisions in many domains. Nevertheless, in realistic scenarios where the consequences of failure are high, running a trained computational agent in the wild poses substantial risk.
\r\n\r\nThe goal of this thesis is to develop and advance techniques that guarantee a learned agent does what we expect it to do. The thesis tackles two central questions:
\r\n\r\n1) Given an agent, how can we predict if it will perform desirably?
\r\n\r\n2) Can we structure the learning process to guarantee desirable post-learning performance?
\r\n\r\nOn the former question, this thesis proposes multiple algorithms to evaluate such agents, finds factors that have high influence on the success of agent evaluation, and open-sources benchmarks for further development in the space.
\r\n\r\nOn the latter question, this thesis formulates desirable agent behavior as a constrained optimization with varying types of constraints depending on the structure afforded to the practitioner. Constraining the search space over the learning process ensures post-learning behaviors will, by definition, perform as desired.
", "doi": "10.7907/n2fg-e554", "publication_date": "2024", "thesis_type": "phd", "thesis_year": "2024" }, { "id": "thesis:16482", "collection": "thesis", "collection_id": "16482", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06032024-152357240", "type": "thesis", "title": "Do Robots Dream of Random Trees? Monte Carlo Tree Search for Dynamical, Partially Observable, and Multi-Agent Systems", "author": [ { "family_name": "Rivi\u00e8re", "given_name": "Benjamin Pierre", "orcid": "0000-0002-0597-5400", "clpid": "Rivi\u00e8re-Benjamin-Pierre" } ], "thesis_advisor": [ { "family_name": "Chung", "given_name": "Soon-Jo", "orcid": "0000-0002-6657-3907", "clpid": "Chung-Soon-Jo" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Chung", "given_name": "Soon-Jo", "orcid": "0000-0002-6657-3907", "clpid": "Chung-Soon-Jo" }, { "family_name": "Hadaegh", "given_name": "Fred", "orcid": "0000-0002-0992-6323", "clpid": "Fred-Hadaegh" }, { "family_name": "Pellegrino", "given_name": "Sergio", "orcid": "0000-0001-9373-3278", "clpid": "Pellegrino-S" } ], "local_group": [ { "literal": "GALCIT" }, { "literal": "div_eng" } ], "abstract": "Autonomous robots are poised to transform various aspects of society, spanning transportation, labor, and scientific space exploration. A critical component to enable their capabilities is the algorithm that interprets sensor data to generate intelligent planned behavior. Although reinforcement learning methods that train parameterized policies offline from data have shown recent success, they are inherently limited when robots inevitably encounter situations outside their training domain. In contrast, optimal control techniques, which compute trajectories in real-time using numerical optimization, typically yield only locally optimal solutions.
\r\n\r\nThis research endeavors to bridge the gap by developing algorithms that compute trajectories in real-time while converging towards globally optimal solutions. Building upon the Monte Carlo Tree Search (MCTS) framework\u2014a stochastic tree search method that simulates future trajectories while balancing exploration and exploitation\u2014the research focus is twofold: (i) constructing an efficient discrete representation of continuous systems in a decision trees, and (ii) searching on the resulting tree while balancing exploration and exploitation to achieve global optimality.
\r\n\r\nThe study spans theoretical analysis, algorithmic design, and hardware demonstrations across dynamical, partially observable, and multi-agent systems. By addressing these critical questions, this research aims to advance the field of autonomous robotics, enabling the deployment of intelligent robots in complex and diverse environments.
", "doi": "10.7907/dbwa-we50", "publication_date": "2024", "thesis_type": "phd", "thesis_year": "2024" }, { "id": "thesis:16297", "collection": "thesis", "collection_id": "16297", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:02182024-070645738", "primary_object_url": { "basename": "changhao_xu_2024_thesis.pdf", "content": "final", "filesize": 61292483, "license": "other", "mime_type": "application/pdf", "url": "/16297/1/changhao_xu_2024_thesis.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Electronic Skin in Robotics and Healthcare: Towards Multimodal Sensing and Intelligent Analysis", "author": [ { "family_name": "Xu", "given_name": "Changhao", "orcid": "0000-0002-6817-3341", "clpid": "Xu-Changhao" } ], "thesis_advisor": [ { "family_name": "Gao", "given_name": "Wei", "orcid": "0000-0002-8503-4562", "clpid": "Gao-Wei" } ], "thesis_committee": [ { "family_name": "Gharib", "given_name": "Morteza", "orcid": "0000-0003-0754-4193", "clpid": "Gharib-M" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Tai", "given_name": "Yu-Chong", "orcid": "0000-0001-8529-106X", "clpid": "Tai-Yu-Chong" }, { "family_name": "Gao", "given_name": "Wei", "orcid": "0000-0002-8503-4562", "clpid": "Gao-Wei" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Skin-interfaced electronics is gradually transforming robotic and medical fields by enabling noninvasive and continuous monitoring of physiological and biochemical information. Despite their promise, current wearable technologies face challenges in several disciplines: Physical sensors are prone to motion-induced noise and lack the capability for effective disease detection, while existing wearable biochemical sensors suffer from operational instability in biofluids, limiting their practicality. Conventional electronic skin contains only a limited category of sensors that are not sufficient for practical applications, and conventional data processing methods for these wearables necessitate manual intervention to filter noise and decipher health-related information.
\r\n\r\nThis thesis presents advances in electronic skin within robotics and healthcare, emphasizing multimodal sensing and data analysis through machine intelligence. Chapter 1 introduces the concept of electronic skin, outlining the emerging sensor technologies and a general machine learning pipeline for data processing. Chapter 2 details the development of multimodal physiological and biochemical sensors that enable long-term continuous monitoring with high sensitivity and stability. Chapter 3 explores the application of integrated electronic skin in robotics, prosthetics, and human machine interactions. Chapter 4 showcases practical implementations of integrated electronic skin with robust sensors for wound monitoring and treatment. Chapter 5 highlights the transformative deployment of artificial intelligence in deconvoluting health profiles on mental health. The last chapter, Chapter 6, delves into the challenges and prospects of artificial intelligence-powered electronic skins, offering predictions for the evolution of smart electronic skins. We envision that multimodal sensing and machine intelligence in electronic skin could significantly advance the field of human machine interactions and personalized healthcare.
", "doi": "10.7907/en0a-ep72", "publication_date": "2024", "thesis_type": "phd", "thesis_year": "2024" }, { "id": "thesis:16363", "collection": "thesis", "collection_id": "16363", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:04292024-172707491", "primary_object_url": { "basename": "Thesis_Cherry_Kevin_Redacted.pdf", "content": "final", "filesize": 20088636, "license": "other", "mime_type": "application/pdf", "url": "/16363/6/Thesis_Cherry_Kevin_Redacted.pdf", "version": "v5.0.0" }, "type": "thesis", "title": "Molecular Pattern Recognition and Supervised Learning in DNA-Based Neural Networks", "author": [ { "family_name": "Cherry", "given_name": "Kevin Matthew", "orcid": "0000-0002-2343-0754", "clpid": "Cherry-Kevin-Matthew" } ], "thesis_advisor": [ { "family_name": "Qian", "given_name": "Lulu", "orcid": "0000-0003-4115-2409", "clpid": "Qian-Lulu" } ], "thesis_committee": [ { "family_name": "Winfree", "given_name": "Erik", "orcid": "0000-0002-5899-7523", "clpid": "Winfree-E" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Rothemund", "given_name": "Paul W. K.", "orcid": "0000-0002-1653-3202", "clpid": "Rothemund-P-W-K" }, { "family_name": "Qian", "given_name": "Lulu", "orcid": "0000-0003-4115-2409", "clpid": "Qian-Lulu" } ], "local_group": [ { "literal": "div_bbe" } ], "abstract": "Adaptation in nature begins at the subcellular, molecular level with the delicate interplay of biomolecule cascades orchestrating the myriad function of cells. The intermingling activity of these cells becomes expressions of complex behavior of multi-cellular system. Nature provides a dazzling array of examples showcasing the variations of intelligent functions. However, in the realm of synthetic construction, what systems have humans managed to engineer, and what are the boundaries of our technological power? In comparison to nature's repertoire, mankind's accomplishments appear rather modest. The intricate behaviors observed in intelligent organisms emerge from the collective interactions and feedback loops among their constituent elements, resulting in the emergence of novel properties and phenomena. To develop large-scale engineered systems exhibiting ever more brain-like, intelligent behaviors, we must first devise new molecular architectures and algorithms designed for adaptation and learning at the molecular scale. My research presented here is a humble step toward those goals. I will present the design of novel molecular systems made from DNA that exhibit complex neural computation and learning behaviors.
\r\n\r\nChapter 2 covers my contribution to scaling up the computing power of DNA circuits. From bacteria following simple chemical gradients to the brain distinguishing complex odor information, the ability to recognize molecular patterns is essential for biological organisms. This type of information-processing function has been implemented using DNA-based neural networks. Winner-take-all computation has been suggested as a potential strategy for enhancing the capability of DNA-based neural networks. Compared to the linear-threshold circuits and Hopfield networks used previously, winner-take-all circuits are computationally more powerful, allow simpler molecular implementation, and are not constrained by coupling the number of patterns and their complexity, so both a large number of simple patterns and a small number of complex patterns can be recognized. Here, we report a systematic implementation of winner-take-all neural networks based on DNA-strand-displacement reactions. We use a previously developed seesaw DNA gate motif, extended to include a simple and robust component that facilitates the cooperative hybridization involved in selecting a \u2018winner.' We show that with this extended seesaw motif, DNA-based neural networks can classify patterns into up to nine categories. Each of these patterns consists of 20 distinct DNA molecules chosen from the set of 100 that represents the 100 bits in 10x10 patterns, with the 20 DNA molecules selected tracing one of the handwritten digits \u20181\u2019 to \u20189.' The network successfully classified test patterns with up to 30 of the 100 bits flipped relative to the digit patterns \u2018remembered\u2019 during training, suggesting that molecular circuits can robustly accomplish the sophisticated task of classifying highly complex and noisy information on the basis of similarity to a memory.
\r\n\r\nChapter 3 investigates the development of a computational neural network model inspired by biological learning mechanisms, particularly focusing on the new mechanisms for learning in a WTA neural network. The study incorporates novel molecular motifs used in inhibited activators and inhibited weights, designed specifically for training from environmental input patterns. These motifs emulate biological systems by facilitating memory storage and retrieval within DNA-based neural networks, similar to synaptic connections and signal processing observed in living organisms. We assess the function of the individual molecular motifs and characterize their specificity in up to 18-species cross-talk experiments. Furthermore, we characterize the network's performance across a wide array of training and test patterns, mirroring the adaptive responses and diverse conditions encountered by biological systems. Additionally, we analyze the computational efficiency and speed of the learning system, comparing it with both the previous non-learning DNA-based WTA model and a direct weight activation model. By exploring the principles of molecular learning, particularly within winner-take-all neural networks, this study aims to advance computational systems by emulating adaptability and resilience observed in biological organisms using robust, new molecular motifs.
", "doi": "10.7907/529f-kf62", "publication_date": "2024", "thesis_type": "phd", "thesis_year": "2024" }, { "id": "thesis:16247", "collection": "thesis", "collection_id": "16247", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:11162023-054627670", "primary_object_url": { "basename": "Jennifer_2023_Caltech_Thesis.pdf", "content": "final", "filesize": 43477334, "license": "other", "mime_type": "application/pdf", "url": "/16247/1/Jennifer_2023_Caltech_Thesis.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "AI for Scientists: Accelerating Discovery Through Knowledge, Data, and Learning", "author": [ { "family_name": "Sun", "given_name": "Jennifer Jianing", "orcid": "0000-0002-0906-6589", "clpid": "Sun-Jennifer-Jianing" } ], "thesis_advisor": [ { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "thesis_committee": [ { "family_name": "Bouman", "given_name": "Katherine L.", "orcid": "0000-0003-0077-4367", "clpid": "Bouman-K-L" }, { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Chaudhuri", "given_name": "Swarat", "orcid": "0000-0002-6859-1391", "clpid": "Chaudhuri-Swarat" }, { "family_name": "Kennedy", "given_name": "Ann", "clpid": "Kennedy-Ann" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "With rapidly growing amounts of experimental data, machine learning is increasingly crucial for automating scientific data analysis. However, many real-world workflows demand expert-in-the-loop attention and require models that not only interface with data, but also with experts and domain knowledge. My research develops full stack solutions that enable scientists to scalably extract insights from diverse and messy experimental data with minimal supervision. My approaches learn from both data and expert knowledge, while exploiting the right level of domain knowledge for generalization. This thesis presents progress towards developing automated scientist-in-the-loop solutions, including methods that automatically discover meaningful structure from data such as self-supervised keypoints from videos of diverse behaving organisms. We will then discuss methods that use these interpretable structures to inject domain knowledge into the learning process, such as guiding representation learning using symbolic programs of behavioral features computed from keypoints. This work is the result of close collaborations with domain experts, such as behavioral neuroscientists, in order to identify bottlenecks and integrate these methods in real-world workflows. My aim is to enable AI that collaborates with scientists to accelerate the scientific process.", "doi": "10.7907/d6y8-4590", "publication_date": "2024", "thesis_type": "phd", "thesis_year": "2024" }, { "id": "thesis:16314", "collection": "thesis", "collection_id": "16314", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:03052024-025023714", "primary_object_url": { "basename": "Thesis-Yujia-Huang.pdf", "content": "final", "filesize": 9659211, "license": "other", "mime_type": "application/pdf", "url": "/16314/1/Thesis-Yujia-Huang.pdf", "version": "v5.0.0" }, "type": "thesis", "title": "Understanding and Improving Reliability of Inference Dynamics in Deep Neural Networks", "author": [ { "family_name": "Huang", "given_name": "Yujia", "orcid": "0000-0001-7667-8342", "clpid": "Huang-Yujia" } ], "thesis_advisor": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "thesis_committee": [ { "family_name": "Bouman", "given_name": "Katherine L.", "orcid": "0000-0003-0077-4367", "clpid": "Bouman-K-L" }, { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" }, { "family_name": "Gkioxari", "given_name": "Georgia", "clpid": "Gkioxari-Georgia" }, { "family_name": "Shi", "given_name": "Yuanyuan", "orcid": "0000-0002-6182-7664", "clpid": "Shi-Yuanyuan" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Reliability is a crucial aspect for the successful deployment of deep learning systems across various domains. In generative modeling, it is essential to create content that adheres to specific rules. In the field of control, ensuring that robots operate safely without falling or entering hazardous areas is paramount. Similarly, in visual perception, the robustness of perception results against perturbations are vital.
\r\n\r\nIn this thesis, we explore the reliability of inference dynamics in deep neural networks such as ResNet, neural Ordinary Differential Equations (ODEs), and diffusion models.\r\nWe begin by examining the inference dynamics in standard networks with a discrete sequence of hidden layers, applying self-consistency and local Lipschitz bounds to enhance robustness against input perturbations. Our exploration then extends to neural ODEs, where the neural network specifies a vector field that continuously transforms the state. We employ forward invariance to achieve robustness, marking the first instance of training neural ODE policies with non-vacuous certified guarantees. The focus shifts next to diffusion models and their inference processes, particularly in adhering to symbolic constraints. For this, we introduce a novel sampling algorithm inspired by stochastic control principles. This algorithm not only guides these models in generating rule-specific content but also sets a new benchmark in symbolic music generation. Our work offers a cohesive understanding of inference dynamics in various deep learning architectures and propose new algorithms to significantly improve their reliability.
", "doi": "10.7907/77cc-6z86", "publication_date": "2024", "thesis_type": "phd", "thesis_year": "2024" }, { "id": "thesis:16177", "collection": "thesis", "collection_id": "16177", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:09062023-095903699", "type": "thesis", "title": "Control of Unknown Dynamical Systems: Robustness and Online Learning of Feedback Control", "author": [ { "family_name": "Ho", "given_name": "Dimitar Mi", "orcid": "0000-0002-7856-985X", "clpid": "Ho-Dimitar-Mi" } ], "thesis_advisor": [ { "family_name": "Doyle", "given_name": "John Comstock", "orcid": "0000-0002-1828-2486", "clpid": "Doyle-J-C" } ], "thesis_committee": [ { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" }, { "family_name": "Doyle", "given_name": "John Comstock", "orcid": "0000-0002-1828-2486", "clpid": "Doyle-J-C" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Over the past few decades, our physical and digital worlds have become increasingly intertwined and reliant on each other. Advancements in areas such as machine learning, online optimization, and control theory, along with ubiquitous access to computational power, have played a crucial role in this technological evolution. As a result, we are now moving towards a future where complex and intelligent dynamical systems, with humans in the loop, govern our daily lives.
\r\n\r\nBuilding advanced control systems is a critical step in this journey, as they enable swift and data-informed decision-making. However, as we aim to create even more sophisticated closed-loop systems, we must proceed with a careful balance of ambition and caution. While the benefits of these interconnected systems are abundant and our dependence on them deepens, ensuring the actual reliability and safety of the systems becomes increasingly challenging due to the growing complexity of their dynamics. This challenge is particularly prominent in safety-critical applications involving physical systems, which often have strict and non-negotiable safety and performance requirements. To establish a harmonious relationship between our physical and digital worlds, it is crucial to develop intelligent closed-loop control systems that are not only fast and efficient, but also reliable and fault-tolerant.
\r\n\r\nThe title of this thesis, \"Control of Unknown Dynamical Systems: Robustness and Online Learning of Feedback Control,\" reflects the central focus of this work on addressing this pressing challenge. The thesis aims to develop theoretical frameworks and tools that provide insights and contribute new approaches to the design of control systems capable of handling the inherent uncertainty in real-world dynamical systems.
\r\n\r\nThe first part of the thesis focuses on the design of closed-loop systems that are robust to dynamic uncertainty, particularly in settings involving nonlinear dynamics and complex control constraints. The second part introduces a general framework for learning-to-control algorithms that provide worst-case guarantees, even in scenarios where the dynamic uncertainty is arbitrarily large. By addressing these key aspects, this work aims to advance our understanding and capabilities in designing control systems that can effectively deal with uncertainty.
", "doi": "10.7907/fb64-vk24", "publication_date": "2024", "thesis_type": "phd", "thesis_year": "2024" }, { "id": "thesis:16093", "collection": "thesis", "collection_id": "16093", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06072023-210232983", "primary_object_url": { "basename": "thesis.pdf", "content": "final", "filesize": 15516327, "license": "other", "mime_type": "application/pdf", "url": "/16093/1/thesis.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Visual and Spatial Representation Learning with Applications in Ecology", "author": [ { "family_name": "Cole", "given_name": "Elijah Henry John", "orcid": "0000000166230966", "clpid": "Cole-Elijah-Henry-John" } ], "thesis_advisor": [ { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" } ], "thesis_committee": [ { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Bouman", "given_name": "Katherine L.", "orcid": "0000-0003-0077-4367", "clpid": "Bouman-K-L" }, { "family_name": "Belongie", "given_name": "Serge J.", "orcid": "0000-0002-0388-5217", "clpid": "Belongie-S-J" } ], "local_group": [ { "literal": "Resnick Sustainability Institute" }, { "literal": "div_eng" } ], "abstract": "Machine learning has the potential to empower scientists, physicians, and other human experts working to solve problems of societal importance. To realize this goal, we need algorithms that can distill useful knowledge from real-world data. However, most machine learning research focuses on benchmarks that seldom reflect real-world challenges, such as learning from limited, noisy, or weak supervision. This thesis develops new benchmarks, algorithms, and problem settings that link fundamental machine learning research to impactful applications in ecology. In Part I, we provide context and motivation for our work. How and why should machine learning researchers work with domain experts on real-world problems? What is the appeal of ecology specifically Part II focuses on visual representation learning with an emphasis on label efficiency. We discuss the strengths and limitations of self-supervised learning, the relationship between concept specificity and representation learning, and multi-label learning with minimal labeled data. Part III covers our work in the emerging field on spatial representation learning. In particular, we consider the problem of modeling the spatial distribution of plant and animal species. We review this important ecological problem from a machine learning perspective before showing how deep learning can transform the way these models are applied (using spatial models to assist image classifiers) and developed (jointly learning spatial distributions and representations). Finally, Part IV concludes and highlights opportunities for future work.", "doi": "10.7907/xp2k-es46", "publication_date": "2023", "thesis_type": "phd", "thesis_year": "2023" }, { "id": "thesis:15101", "collection": "thesis", "collection_id": "15101", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:02072023-112938936", "type": "thesis", "title": "The Identification of Discrete Mixture Models", "author": [ { "family_name": "Gordon", "given_name": "Spencer Lane", "orcid": "0000-0002-7101-2370", "clpid": "Gordon-Spencer-Lane" } ], "thesis_advisor": [ { "family_name": "Vidick", "given_name": "Thomas G.", "orcid": "0000-0002-6405-365X", "clpid": "Vidick-T" }, { "family_name": "Schulman", "given_name": "Leonard J.", "orcid": "0000-0001-9901-2797", "clpid": "Schulman-L-J" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Vidick", "given_name": "Thomas G.", "orcid": "0000-0002-6405-365X", "clpid": "Vidick-T" }, { "family_name": "Schulman", "given_name": "Leonard J.", "orcid": "0000-0001-9901-2797", "clpid": "Schulman-L-J" }, { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "In this thesis we discuss a variety of results on learning and identifying discrete mixture models, i.e., distributions that are a convex combination of k from a known class C of distributions. We first consider the case where C is the class of binomial distributions, before generalizing to the case of product distributions. We provide a necessary condition for identifiability of mixture of products distributions as well as a generalization to structured mixtures over multiple latent variables.", "doi": "10.7907/ebf5-0b48", "publication_date": "2023", "thesis_type": "phd", "thesis_year": "2023" }, { "id": "thesis:14980", "collection": "thesis", "collection_id": "14980", "cite_using_url": "https://resolver.caltech.edu/CaltechThesis:07202022-040725024", "primary_object_url": { "basename": "Thesis_Tongxin-Li.pdf", "content": "final", "filesize": 12516785, "license": "other", "mime_type": "application/pdf", "url": "/14980/1/Thesis_Tongxin-Li.pdf", "version": "v5.0.0" }, "type": "thesis", "title": "Learning-Augmented Control and Decision-Making: Theory and Applications in Smart Grids", "author": [ { "family_name": "Li", "given_name": "Tongxin", "orcid": "0000-0002-9806-8964", "clpid": "Li-Tongxin" } ], "thesis_advisor": [ { "family_name": "Low", "given_name": "Steven H.", "orcid": "0000-0001-6476-3048", "clpid": "Low-S-H" }, { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Low", "given_name": "Steven H.", "orcid": "0000-0001-6476-3048", "clpid": "Low-S-H" }, { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" }, { "family_name": "Mazumdar", "given_name": "Eric V.", "orcid": "0000-0002-1815-269X", "clpid": "Mazumdar-Eric" } ], "local_group": [ { "literal": "Resnick Sustainability Institute" }, { "literal": "div_eng" } ], "abstract": "Achieving carbon neutrality by 2050 does not only lead to the increasing penetration of renewable energy, but also an explosive growth of smart meter data. Recently, augmenting classical methods in real-world cyber-physical systems such as smart grids with black-box AI tools, forecasts, and ML algorithms has attracted a lot of growing interest. Integrating AI techniques into smart grids, on the one hand, provides a new approach to handle the uncertainties caused by renewable resources and human behaviors, but on the other hand, creates practical issues such as reliability, stability, privacy, and scalability, etc. to the AI-integrated algorithms.
\r\n\r\nThis dissertation focuses on solving problems raised in designing learning-augmented control and decision-making algorithms.
\r\n \r\nThe results presented in this dissertation are three-fold. We first study a problem in linear quadratic control, where imperfect/untrusted AI predictions of system perturbations are available. We show that it is possible to design a learning-augmented algorithm with performance guarantees that is aggressive if the predictions are accurate and conservative if they are imperfect. Machine-learned black-box policies are ubiquitous for nonlinear control problems. Meanwhile, crude model information is often available for these problems from, e.g., linear approximations of nonlinear dynamics. We next study the problem of equipping a black-box control policy with model-based advice for nonlinear control on a single trajectory. We first show a general negative result that a naive convex combination of a black-box policy and a linear model-based policy can lead to instability, even if the two policies are both stabilizing. We then propose an adaptive \u03bb-confident policy, with a coefficient \u03bb indicating the confidence in a black-box policy, and prove its stability. With bounded nonlinearity, in addition, we show that the adaptive \u03bb-confident policy achieves a bounded competitive ratio when a black-box policy is near-optimal. Finally, we propose an online learning approach to implement the adaptive \u03bb-confident policy and verify its efficacy in case studies about the Cart-Pole problem and a real-world electric vehicle (EV) charging problem with data bias due to COVID-19.
\r\n\r\nAggregators have emerged as crucial tools for the coordination of distributed, controllable loads. To be used effectively, an aggregator must be able to communicate the available flexibility of the loads they control, known as the aggregate flexibility to a system operator. However, most existing aggregate flexibility measures often are slow-timescale estimations and much less attention has been paid to real-time coordination between an aggregator and an operator. In the second part of this dissertation, we consider solving an online decision-making problem in a closed-loop system and present a design of real-time aggregate flexibility feedback, termed the maximum entropy feedback (MEF). In addition to deriving analytic properties of the MEF, combining learning and control, we show that it can be approximated using reinforcement learning and used as a penalty term in a novel control algorithm--the penalized predictive control (PPC) that enables efficient communication, fast computation, and lower costs. We illustrate the efficacy of the PPC using a dataset from an adaptive electric vehicle charging network and show that PPC outperforms classical MPC. We show that under certain regularity assumptions, the PPC is optimal. We illustrate the efficacy of the PPC using a dataset from an adaptive electric vehicle charging network and show that PPC outperforms classical model predictive control (MPC). In a theoretical perspective, a two-controller problem is formulated. A central controller chooses an action from a feasible set that is determined by time-varying and coupling constraints, which depend on all past actions and states. The central controller's goal is to minimize the cumulative cost; however, the controller has access to neither the feasible set nor the dynamics directly, which are determined by a remote local controller. Instead, the central controller receives only an aggregate summary of the feasibility information from the local controller, which does not know the system costs. We show that it is possible for an online algorithm using feasibility information to nearly match the dynamic regret of an online algorithm using perfect information whenever the feasible sets satisfy some criterion, which is satisfied by inventory and tracking constraints.
\r\n\r\nThe third part of this dissertation consists of examples of learning, inference, and data analysis methods for power system identification and electric charging. We present a power system identification problem with noisy nodal measurements and efficient algorithms, based on fundamental trade-offs between the number of measurements, the complexity of the graph class, and the probability of error. Next, we specifically consider prediction and unsupervised learning tasks in EV charging. We provide basic data analysis results of a public dataset released by Caltech and develop a novel iterative clustering method for classifying time series of EV charging rates.
", "doi": "10.7907/cdf6-0w78", "publication_date": "2023", "thesis_type": "phd", "thesis_year": "2023" }, { "id": "thesis:15130", "collection": "thesis", "collection_id": "15130", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:04092023-200347393", "primary_object_url": { "basename": "Thesis_Charles-Guan.pdf", "content": "final", "filesize": 9022504, "license": "other", "mime_type": "application/pdf", "url": "/15130/2/Thesis_Charles-Guan.pdf", "version": "v7.0.0" }, "type": "thesis", "title": "Neural Coding of Finger Movements in Human Posterior Parietal Cortex and Motor Cortex", "author": [ { "family_name": "Guan", "given_name": "Charles", "orcid": "0000-0002-8040-8844", "clpid": "Guan-Charles" } ], "thesis_advisor": [ { "family_name": "Andersen", "given_name": "Richard A.", "orcid": "0000-0002-7947-0472", "clpid": "Andersen-R-A" } ], "thesis_committee": [ { "family_name": "Meister", "given_name": "Markus", "orcid": "0000-0003-2136-6506", "clpid": "Meister-M" }, { "family_name": "Rutishauser", "given_name": "Ueli", "orcid": "0000-0002-9207-7069", "clpid": "Rutishauser-U" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Andersen", "given_name": "Richard A.", "orcid": "0000-0002-7947-0472", "clpid": "Andersen-R-A" } ], "local_group": [ { "literal": "div_bbe" } ], "abstract": "We use our hands constantly in our everyday lives. This seemingly simple ability is disrupted in individuals with cervical spinal cord injuries. By circumventing injured signal pathways, brain-computer interfaces (BCIs) promise to enable such individuals to control artificial limbs for everyday use. However, existing BCI limb control remains coarse and inflexible, because we do not understand how the recorded neural activity relates to dexterous movement. As a result, BCI control in physical settings remains frustratingly difficult for paralyzed users. To improve dexterous BCI control, I studied the neural coding of individual finger movements in the posterior parietal cortex and motor cortex of tetraplegic participants. These regions are directly involved in dexterous hand movements and are candidates for BCI recording implants. Finger coding matched the correlation structure and dynamics of able-bodied usage, reflecting preserved motor circuits even after paralysis. Individual finger movements of each hand were coded in a factorized, correlated manner that still allowed decoding. Participants controlled artificial fingers with state-of-the-art accuracy. Finally, we studied the temporal dynamics of neural control to understand how existing models of neural activity extend to BCI control. These findings contribute to the understanding of human hand movements and advance the development of dexterous BCIs.", "doi": "10.7907/31rt-cy14", "publication_date": "2023", "thesis_type": "phd", "thesis_year": "2023" }, { "id": "thesis:14556", "collection": "thesis", "collection_id": "14556", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:04242022-005200355", "primary_object_url": { "basename": "Thesis_Final.pdf", "content": "final", "filesize": 194962657, "license": "other", "mime_type": "application/pdf", "url": "/14556/1/Thesis_Final.pdf", "version": "v6.0.0" }, "type": "thesis", "title": "Where the Wild Things Are: Computer Vision for Global-Scale Biodiversity Monitoring", "author": [ { "family_name": "Beery", "given_name": "Sara Meghan", "orcid": "0000-0002-2544-1844", "clpid": "Beery-Sara-Meghan" } ], "thesis_advisor": [ { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Bouman", "given_name": "Katherine L.", "orcid": "0000-0003-0077-4367", "clpid": "Bouman-K-L" }, { "family_name": "Belongie", "given_name": "Serge J.", "orcid": "0000-0002-0388-5217", "clpid": "Belongie-S-J" }, { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" } ], "local_group": [ { "literal": "Resnick Sustainability Institute" }, { "literal": "div_eng" } ], "abstract": "We require a real-time, modular earth observation system that unites efforts across research groups in order to provide the necessary information necessary for global-scale impact in sustainability and conservation in the face of climate change. The development of such systems requires collaborative, interdisciplinary approaches that translate diverse sources of raw information into accessible scientific insight. For example, we need to monitor species in real time and in greater detail to quickly understand which conservation efforts are most effective and take corrective action. Current ecological monitoring systems generate data far faster than researchers can analyze it, making scaling up impossible without automated data processing. However, ecological data collected in the field presents a number of challenges that current methods, like deep learning, are not designed to tackle. These include strong spatiotemporal correlations, imperfect data quality, fine-grained categories, and long-tailed distributions. Our work seeks to overcome these challenges, and this thesis includes methods which can learn from imperfect data, systematic frameworks and benchmarks for measuring and overcoming performance drops due to domain shift, and the development and deployment of efficient human-AI systems that have made significant real-world conservation impact.", "doi": "10.7907/m4mt-2q51", "publication_date": "2023", "thesis_type": "phd", "thesis_year": "2023" }, { "id": "thesis:14994", "collection": "thesis", "collection_id": "14994", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:08052022-231458463", "type": "thesis", "title": "Reliable Learning and Control in Dynamic Environments: Towards Unified Theory and Learned Robotic Agility", "author": [ { "family_name": "Shi", "given_name": "Guanya", "orcid": "0000-0002-9075-3705", "clpid": "Shi-Guanya" } ], "thesis_advisor": [ { "family_name": "Chung", "given_name": "Soon-Jo", "orcid": "0000-0002-6657-3907", "clpid": "Chung-Soon-Jo" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "thesis_committee": [ { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" }, { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" }, { "family_name": "Chung", "given_name": "Soon-Jo", "orcid": "0000-0002-6657-3907", "clpid": "Chung-Soon-Jo" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Recent breathtaking advances in machine learning beckon to their applications in a wide range of real-world autonomous systems. However, for safety-critical settings such as agile robotic control in hazardous environments, we must confront several key challenges before widespread deployment. Most importantly, the learning system must interact with the rest of the autonomous system (e.g., highly nonlinear and non-stationary dynamics) in a way that safeguards against catastrophic failures with formal guarantees. In addition, from both computational and statistical standpoints, the learning system must incorporate prior knowledge for efficiency and generalizability.
\r\n\r\nThis thesis presents progress towards establishing a unified framework that fundamentally connects learning and control. First, Part I motivates the benefit and necessity of such a unified framework by the Neural-Control Family, a family of nonlinear deep-learning-based control methods with not only stability and robustness guarantees but also new capabilities in agile robotic control. Then Part II discusses three unifying interfaces between learning and control: (1) online meta-adaptive control, (2) competitive online optimization and control, and (3) online learning perspectives on model predictive control. All interfaces yield settings that jointly admit both learning-theoretic and control-theoretic guarantees.
", "doi": "10.7907/8rz4-7b35", "publication_date": "2023", "thesis_type": "phd", "thesis_year": "2023" }, { "id": "thesis:15149", "collection": "thesis", "collection_id": "15149", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:04292023-003436131", "type": "thesis", "title": "Enabling Robust and User-Customized Bipedal Locomotion on Lower-Body Assistive Devices via Hybrid System Theory and Preference-Based Learning", "author": [ { "family_name": "Tucker", "given_name": "Maegan Lindsay", "orcid": "0000-0001-7363-6809", "clpid": "Tucker-Maegan-Lindsay" } ], "thesis_advisor": [ { "family_name": "Ames", "given_name": "Aaron D.", "orcid": "0000-0003-0848-3177", "clpid": "Ames-A-D" } ], "thesis_committee": [ { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" }, { "family_name": "Ames", "given_name": "Aaron D.", "orcid": "0000-0003-0848-3177", "clpid": "Ames-A-D" }, { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Practical robotic assistive devices have the potential to transform many aspects of our society, from enabling locomotive autonomy to facilitating rehabilitation. However, as is typically the case when having autonomous systems interact closely with humans, one must simultaneously solve multiple grand challenges. My work focuses specifically on 1) leveraging hybrid system theory to achieve stable and robust walking that generalizes well across various human models and environmental conditions, and 2) developing an online learning strategy to customize the experimental walking for individual user comfort. The presented methodology is grounded in realizing lower-body exoskeleton locomotion for subjects with motor complete paraplegia, with extensions to other robotic applications. The contributions are broken down as follows.
\r\n\r\nFirst, by leveraging tools from nonlinear control theory, I propose techniques for systematically addressing locomotive robustness. These techniques include: using saltation matrices to generate robust gaits with experimental demonstrations on the Atalante lower-body exoskeleton; and developing an input-to-state stability perspective to certify robustness to uncertain impact events. Importantly, these methods aim to better understand the mathematical conditions underlying robust locomotion -- a necessary step towards realizing safe locomotion across varying human models and environmental conditions. Second, I develop a preference-based learning framework to explicitly optimize user comfort during exoskeleton locomotion (achieved using the aforementioned nonlinear control methodology) by learning directly from subjective feedback. This framework is implemented in real-world settings, including the clinical realization of user-preferred locomotion for two subjects with motor complete paraplegia.Third, the extensibility of this framework is demonstrated through three general robotic applications: tuning constraints of the gait generation optimization problem with demonstrations on a planar biped; tuning Lyapunov-based controller gains on a 3D biped; and tuning control barrier function parameters for performant yet safe exploration on a quadrupedal platform. Lastly, I discuss other relevant clinical considerations for lower-body assistive devices including how exoskeleton locomotion influences metabolic cost of transport, the study of latent factors underlying user-preferred walking, and embedding musculoskeletal models directly in the gait generation process.
", "doi": "10.7907/j9hk-xa17", "publication_date": "2023", "thesis_type": "phd", "thesis_year": "2023" }, { "id": "thesis:15166", "collection": "thesis", "collection_id": "15166", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:05122023-085319625", "type": "thesis", "title": "Grasp, Speech, and Internal Speech Representation in the Human Cortical Grasp Circuit", "author": [ { "family_name": "Wandelt", "given_name": "Sarah Kim", "orcid": "0000-0001-9551-8491", "clpid": "Wandelt-Sarah-Kim" } ], "thesis_advisor": [ { "family_name": "Andersen", "given_name": "Richard A.", "orcid": "0000-0002-7947-0472", "clpid": "Andersen-R-A" } ], "thesis_committee": [ { "family_name": "Adolphs", "given_name": "Ralph", "orcid": "0000-0002-8053-9692", "clpid": "Adolphs-R" }, { "family_name": "Rutishauser", "given_name": "Ueli", "orcid": "0000-0002-9207-7069", "clpid": "Rutishauser-U" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Andersen", "given_name": "Richard A.", "orcid": "0000-0002-7947-0472", "clpid": "Andersen-R-A" } ], "local_group": [ { "literal": "div_bbe" } ], "abstract": "The ability to move freely and to connect with others through communication is invaluable for human independence. In this thesis, we explore how brain-machine interfaces (BMIs) can help patients affected by movement or speech deficits to recover lost human experiences. This work builds on previous findings indicating that premotor and posterior parietal areas are involved in movement generation and language processes. These higher-level brain areas do not only engage in movement execution, but also during planning, representing rich behavioral patterns that can be leveraged for BMI applications. In this work, we investigated how the ventral premotor cortex (PMv), the posterior parietal cortex (PPC), and the sensorimotor cortex (S1) represent grasp and speech processes at a single-neuron level. Using multielectrode Utah arrays, neuronal populations were recorded in tetraplegic human participants. We found that the supramarginal gyrus (SMG), PMv and S1 significantly encode motor imagery of grasping. By studying the cognitive processes underlying neural activity during the cue phase of grasping, we found a transition from cue-modality dependent to cue-modality independent grasp representation in SMG, the anterior intraparietal cortex (AIP) and PMv. Our findings suggest SMG integrates audio, written, and image cue modalities, but more similarly represents audio and written cues, indicating language involvement. We confirmed this hypothesis by demonstrating that SMG encodes spoken words, engaging different motor plans for speech compared to grasping even when the semantic content remained unchanged. These results suggest a BMI could be trained to decode both grasp motor imagery and speech from one brain area. Lastly, we showed that SMG is highly involved in language processes, modulating for written word recognition, auditory tones, vocalized speech, and internal speech. As a proof-of-concept, we built a real-time internal speech BMI from signals recorded in SMG that can decode eight words with high accuracy. This work is the first of its kind, demonstrating internal speech can be robustly decoded from an implant in a single brain area. We find high neural SMG generalization between seeing a written word, saying it internally and vocalizing the word, suggesting shared cognitive functions between different language processes. Furthermore, words in different languages are represented in SMG. This thesis advances the BMI field by providing a better understanding of the neural processes that underly grasp motor imagery and language. To summarize, our findings suggest that studying higher-level brain areas can lead to the development of more effective and versatile brain-machine interfaces.", "doi": "10.7907/0461-g107", "publication_date": "2023", "thesis_type": "phd", "thesis_year": "2023" }, { "id": "thesis:15041", "collection": "thesis", "collection_id": "15041", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:10132022-000100592", "primary_object_url": { "basename": "bernstein_thesis.pdf", "content": "final", "filesize": 2113002, "license": "other", "mime_type": "application/pdf", "url": "/15041/1/bernstein_thesis.pdf", "version": "v5.0.0" }, "type": "thesis", "title": "Optimisation & Generalisation in Networks of Neurons", "author": [ { "family_name": "Bernstein", "given_name": "Jeremy David", "orcid": "0000-0001-9110-7476", "clpid": "Bernstein-Jeremy-David" } ], "thesis_advisor": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "thesis_committee": [ { "family_name": "Tropp", "given_name": "Joel A.", "orcid": "0000-0003-1024-1791", "clpid": "Tropp-J-A" }, { "family_name": "Liu", "given_name": "Ming-Yu", "orcid": "0000-0002-2951-2398", "clpid": "Liu-Ming-Yu" }, { "family_name": "Meister", "given_name": "Markus", "orcid": "0000-0003-2136-6506", "clpid": "Meister-M" }, { "family_name": "Thomson", "given_name": "Matthew", "orcid": "0000-0003-1021-1234", "clpid": "Thomson-M-W" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_bbe" } ], "abstract": "The goal of this thesis is to develop the optimisation and generalisation theoretic foundations of learning in artificial neural networks. The thesis tackles two central questions. Given training data and a network architecture:
\r\n\r\nOn optimisation, an essential feature of neural network training is that the network weights affect the loss function only indirectly through their appearance in the network architecture. This thesis proposes a three-step framework for deriving novel \u201carchitecture aware\u201d optimisation algorithms. The first step\u2014termed functional majorisation\u2014is to majorise a series expansion of the loss function in terms of functional perturbations. The second step is to derive architectural perturbation bounds that relate the size of functional perturbations to the size of weight perturbations. The third step is to substitute these architectural perturbation bounds into the functional majorisation of the loss and to obtain an optimisation algorithm via minimisation. This constitutes an application of the majorise-minimise meta-algorithm to neural networks.
\r\n\r\nOn generalisation, a promising recent line of work has applied PAC-Bayes theory to derive non-vacuous generalisation guarantees for neural networks. Since these guarantees control the average risk of ensembles of networks, they do not address which individual network should generalise best. To close this gap, the thesis rekindles an old idea from the kernels literature: the Bayes point machine. A Bayes point machine is a single classifier that approximates the aggregate prediction of an ensemble of classifiers. Since aggregation reduces the variance of ensemble predictions, Bayes point machines tend to generalise better than other ensemble members. The thesis shows that the space of neural networks consistent with a training set concentrates on a Bayes point machine if both the network width and normalised margin are sent to infinity. This motivates the practice of returning a wide network of large normalised margin.
\r\n\r\nPotential applications of these ideas include novel methods for uncertainty quantification, more efficient numerical representations for neural hardware, and optimisers that transfer hyperparameters across learning problems.
", "doi": "10.7907/1jz8-5t85", "publication_date": "2023", "thesis_type": "phd", "thesis_year": "2023" }, { "id": "thesis:15260", "collection": "thesis", "collection_id": "15260", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06012023-194820518", "primary_object_url": { "basename": "Thesis-Edit.pdf", "content": "final", "filesize": 18957251, "license": "other", "mime_type": "application/pdf", "url": "/15260/1/Thesis-Edit.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Geometry and Dynamical Systems in Machine Learning and Control", "author": [ { "family_name": "Dorobantu", "given_name": "Victor David", "orcid": "0000-0002-2797-7802", "clpid": "Dorobantu-Victor-David" } ], "thesis_advisor": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "thesis_committee": [ { "family_name": "Dabiri", "given_name": "John O.", "orcid": "0000-0002-6722-9008", "clpid": "Dabiri-J-O" }, { "family_name": "Schroeder", "given_name": "Peter", "orcid": "0000-0002-0323-7674", "clpid": "Schr\u00f6der-P" }, { "family_name": "Azizzadenesheli", "given_name": "Kamyar", "orcid": "0000-0001-8507-1868", "clpid": "Azizzadenesheli-Kamyar" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "For many problems of interest in machine learning and control, we have access to rich information about underlying geometry and dynamics; we can leverage this information to build robust and performant solutions in new algorithms, optimizations, and designs. In this thesis we study four problem settings to stress this central assumption. First, we study conformal generative modeling, using computational geometry techniques to simplify and register complex 2D surfaces and enabling the use of a variety of flow-based generative models as plug-and-play subroutines. Second, we study data-driven robust optimization problems in control, modeling the precise impact of dynamics uncertainty in several control frameworks using convex geometry. Third, we study compactly-restrictable policy optimization, constraining the available states and actions in reinforcement learning and optimal control problems to be consistent with the inherent dynamics of the systems to be controlled. Finally, we study nonlinear model predictive control on Lie groups as applied to a 3D hopping robot platform, developing a control methodology compatible with nontrivial state space geometry and hybrid system dynamics.", "doi": "10.7907/x271-r088", "publication_date": "2023", "thesis_type": "phd", "thesis_year": "2023" }, { "id": "thesis:15264", "collection": "thesis", "collection_id": "15264", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06012023-213052258", "type": "thesis", "title": "Machine Learning and Data Assimilation for Blending Incomplete Models and Noisy Data", "author": [ { "family_name": "Levine", "given_name": "Matthew Emanuel", "orcid": "0000-0002-5627-3169", "clpid": "Levine-Matthew-Emanuel" } ], "thesis_advisor": [ { "family_name": "Stuart", "given_name": "Andrew M.", "orcid": "0000-0001-9091-7266", "clpid": "Stuart-A-M" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Owhadi", "given_name": "Houman", "orcid": "0000-0002-5677-1600", "clpid": "Owhadi-H" }, { "family_name": "Bouman", "given_name": "Katherine L.", "orcid": "0000-0003-0077-4367", "clpid": "Bouman-K-L" }, { "family_name": "Stuart", "given_name": "Andrew M.", "orcid": "0000-0001-9091-7266", "clpid": "Stuart-A-M" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "The prediction and inference of dynamical systems is of widespread interest across scientific and engineering disciplines. Data assimilation (DA) offers a well-established and successful paradigm for blending such models with noisy observational data. However, traditional DA-based inference often fails when available data are insufficiently informative. Chapter 2 copes with this challenge by introducing constraints into Ensemble Kalman Filtering, which is shown to improve forecasting of glucose dynamics in real patient-level clinical data. Chapter 3 addresses this identifiability challenge by instead developing a simplified, reduced-order stochastic model for glucose dynamics that is more easily identified from patient data. Despite these successes, the forecasting performance of the methods are fundamentally limited by the fidelity of the employed model, which is often not fully understood a priori.
\r\n\r\nChapter 4 presents a general picture of how noisy, partially-observed time-series data can be used to learn flexible (e.g., neural network-based) corrections to a pre-specified mechanistic model. In Chapter 5, the proposed methodology is then validated in simulated settings for glucose-insulin models. Chapter 6 provides further perspective on learning flexible model corrections, comparing approaches that use i) gradient-based or gradient-free optimization, ii) temporal or time-averaged data, iii) different model parameterizations, iv) deterministic and stochastic corrections, and v) physical conservation laws to constrain inference.
\r\n\r\nChapter 7 studies how these perspectives on machine learning and dynamical systems can help us understand the roles of biochemical networks. In particular, it considers protein dimerization networks from the lens of approximation theory and evaluates how the equilibria of these networks can be fine-tuned to perform a variety of biological computations.
", "doi": "10.7907/b82h-ye78", "publication_date": "2023", "thesis_type": "phd", "thesis_year": "2023" }, { "id": "thesis:15275", "collection": "thesis", "collection_id": "15275", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06022023-032907616", "primary_object_url": { "basename": "Taylor_Thesis_Submitted.pdf", "content": "final", "filesize": 31746445, "license": "other", "mime_type": "application/pdf", "url": "/15275/1/Taylor_Thesis_Submitted.pdf", "version": "v5.0.0" }, "type": "thesis", "title": "Robust Safety-Critical Control: A Lyapunov and Barrier Approach", "author": [ { "family_name": "Taylor", "given_name": "Andrew James", "orcid": "0000-0002-5990-590X", "clpid": "Taylor-Andrew-James" } ], "thesis_advisor": [ { "family_name": "Ames", "given_name": "Aaron D.", "orcid": "0000-0003-0848-3177", "clpid": "Ames-A-D" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" }, { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" }, { "family_name": "Ames", "given_name": "Aaron D.", "orcid": "0000-0003-0848-3177", "clpid": "Ames-A-D" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Accompanying the technological advances of the past decade has been the promise for widespread growth of autonomous systems into nearly all domains of human society, including manufacturing, transportation, and healthcare. At the same time, there have been several tragic failures that reveal potential risks with the expansion of autonomous systems into everyday life, and indicate that it is vital for safety to be accounted for in the design of control systems.
\r\n\r\nThis thesis seeks to develop a theory of robust safety-critical control for autonomous systems. This theory will be built upon the foundational tools of Control Lyapunov Functions (CLFs) and Control Barrier Functions (CBFs), which provide a powerful paradigm for the design of model-based safety-critical controllers. The dependence of CLF and CBF-based controllers on a system model makes them susceptible to modeling inaccuracies, potentially resulting in unsafe behavior when deploying these controllers on real-world systems.
\r\n\r\nIn this thesis I present methods for resolving four classes of model inaccuracies referred to as model error, disturbances, measurement error, and input sampling, which are commonly faced challenges when designing controllers for robotic systems. The proposed methods are unified by their shared use of CLFs and CBFs to produce controllers possessing rigorous and robust safety guarantees that can be demonstrated in simulation or experimentally. A hallmark of these methods is a focus on enabling control synthesis through convex optimization, which ensures that controllers can be efficiently computed on real-world robotic hardware platforms.
\r\n\r\nIn addressing model error, I consider both data-driven learning approaches and adaptive control approaches. I present three episodic learning frameworks that iteratively augment existing CLF and CBF-based controllers specified via convex optimization problems to improve the stability and safety properties of a system, which I demonstrate in simulation and experimentally. I also establish a relationship between the degradation of stability and safety properties with the magnitude of residual learning error through the perspective of Input-to-State Stability (ISS) and Input-to-State Safety (ISSf). Lastly, I develop an adaptive safety-critical control framework for systems with parametric model error through the notion of adaptive CBFs.
\r\n\r\nIn addressing disturbances, I resolve challenges in balancing performance and robustness with ISSf-based controllers through the notion of Tunable Input-to-State Safety (TISSf), which permits prioritizing robustness to disturbances only when safety requirements are close to being violated. I demonstrate the capabilities of TISSf-based control design experimentally on an autonomous semi-trailer truck system that is subject to input disturbances due to complex unmodeled actuator dynamics. Lastly, I develop a framework for achieving ISSf-like finite-time safety guarantees for discrete-time systems subject to stochastic disturbances through the use of CBFs and convex optimization.
\r\n\r\nIn addressing measurement error, I develop the notion of Measurement-Robust CBFs (MR-CBFs), which permit control synthesis through convex optimization in the presence of imperfect measurements. I demonstrate the capability of MR-CBFs on an experimental Segway system using a vision-based measurement system, validating the tractability of using controllers specified through increasingly complex classes of convex optimization problems on real-world systems. Lastly, I present an application of Preference Based Learning (PBL) in tuning the robustness parameters of a CBF-based controller, demonstrating the first use of PBL with CBFs and providing a tool for tuning the safety and performance of the robust controllers proposed in this thesis.
\r\n\r\nIn addressing input sampling, I consider both sampled-data and event-triggered paradigms for modeling input sampling. I provide a method for synthesizing CLF-based controllers for sampled-data systems by integrating feedback linearization with approximate discrete-time models, leading to a significant improvement over continuous-time CLF-based controllers implemented with input sampling. I then develop a framework for achieving safety of sampled-data systems through approximate discrete-time models through the notion of practical safety and Sampled-Data CBFs (SD-CBFs), which I demonstrate with convex-optimization based controllers in simulation. Lastly, I develop a method for event-triggered safety-critical control that uses ISSf to achieve safety while satisfying the requirement of a minimum interevent time.
\r\n\r\nCollectively, these contributions constitute a significant advance in the theory of robust safety-critical control by establishing a framework, unified by the use of CLFs and CBFs in conjunction with convex optimization, that addresses a wide class of challenges faced in the design of safety-critical control systems.
", "doi": "10.7907/bpht-by81", "publication_date": "2023", "thesis_type": "phd", "thesis_year": "2023" }, { "id": "thesis:16091", "collection": "thesis", "collection_id": "16091", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06072023-134620248", "primary_object_url": { "basename": "o'connell_michael_2023.pdf", "content": "final", "filesize": 20683543, "license": "other", "mime_type": "application/pdf", "url": "/16091/1/o'connell_michael_2023.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Methods for Robust Learning-Based Control", "author": [ { "family_name": "O'Connell", "given_name": "Michael Thomas", "orcid": "0000-0001-6681-8823", "clpid": "O'Connell-Michael-Thomas" } ], "thesis_advisor": [ { "family_name": "Chung", "given_name": "Soon-Jo", "orcid": "0000-0002-6657-3907", "clpid": "Chung-Soon-Jo" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" }, { "family_name": "Pellegrino", "given_name": "Sergio", "orcid": "0000-0001-9373-3278", "clpid": "Pellegrino-S" }, { "family_name": "Chung", "given_name": "Soon-Jo", "orcid": "0000-0002-6657-3907", "clpid": "Chung-Soon-Jo" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "This thesis addresses the general problem of improving control, safety, and reliability of multi-rotor drones in various challenging conditions by introducing novel deep-learning-based approaches. These approaches are designed to tackle specific issues that multi-rotor drones face during operation, such as near-ground trajectory control, high-speed wind disturbances, actuation delays, and motor failures. The thesis is organized into four main chapters, plus an introduction and conclusion. Each of the main chapters focuses on a unique approach to address a particular challenge of deep-learning-based control methods. Chapter 2 presents Neural-Lander, a deep-learning-based robust nonlinear controller that significantly improves quadrotor control performance during landing by accounting for complex aerodynamic effects. This chapter addresses key challenges to incorporating learned residual dynamics into a control architecture, laying the groundwork for the subsequent chapters. Chapters 3 and 4 introduce Neural-Fly, a learning-based approach that uses Domain Adversarially Invariant Meta-Learning (DAIML) and adaptive control to enable rapid online learning and precise flight control under a wide range of wind conditions. Chapter 5 proposes a lightweight augmentation method that enhances trajectory tracking performance for UAVs by effectively compensating for motor dynamics and digital transport delays. This method is extensible to a range of control methods, including learning-based approaches. Chapter 6 explores a novel sparse failure identification method for detecting and compensating for motor failures in over-actuated UAVs, contributing to the development of robust fault detection and compensation strategies for a safer and more reliable operation. This method builds on the Neural-Fly online learning framework and extends it to handle a wider range of conditions, including complete actuator failures. Together, these chapters address key challenges in safe and reliable learning-based control and demonstrate the potential of deep-learning-based control methods.
", "doi": "10.7907/2xnc-t162", "publication_date": "2023", "thesis_type": "phd", "thesis_year": "2023" }, { "id": "thesis:16114", "collection": "thesis", "collection_id": "16114", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06122023-232029846", "primary_object_url": { "basename": "Werner_Thesis.pdf", "content": "final", "filesize": 34782096, "license": "other", "mime_type": "application/pdf", "url": "/16114/2/Werner_Thesis.pdf", "version": "v6.0.0" }, "type": "thesis", "title": "Uncertainty and Decentralization: Two Themes in an Energy Transformation", "author": [ { "family_name": "Werner", "given_name": "Lucien Desloge", "orcid": "0000-0003-1613-1702", "clpid": "Werner-Lucien-Desloge" } ], "thesis_advisor": [ { "family_name": "Low", "given_name": "Steven H.", "orcid": "0000-0001-6476-3048", "clpid": "Low-S-H" }, { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" }, { "family_name": "Low", "given_name": "Steven H.", "orcid": "0000-0001-6476-3048", "clpid": "Low-S-H" }, { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" } ], "local_group": [ { "literal": "Resnick Sustainability Institute" }, { "literal": "div_eng" } ], "abstract": "Over the last two decades, the rapidly decreasing units costs of solar, wind, and energy storage technologies have launched a fundamental transformation in how electric power is produced, distributed, and consumed. Proliferation of these technologies has effected a shift towards a more decentralized, flexible, and sustainable energy system that can meet the growing demand for energy while reducing greenhouse gas emissions from fossil fuels. The work in this thesis studies two principal themes in this transformation: uncertainty and decentralization.
\r\n\r\nUncertainty is a key challenge in the modern grid resulting from the weather dependence of variable renewables and volatile loads like electric vehicles distributed throughout the grid. Electricity markets, whose function is to regulate the precise balance of supply and demand across the system, face a pressing need for dispatch mechanisms that account for uncertainty while providing participation incentives for generators and loads. We introduce a framework for multi-stage market dispatch and pricing under a general description of forecast uncertainty that enables system operators to explicitly incorporate uncertainty into market-clearing prices. In related work, we study mechanisms that guarantee feasibility of multi-interval dispatch under robust uncertainty and provide participation incentives for shiftable demand response in forward multi-interval markets.
\r\n\r\nThe trend towards a more decentralized energy system stems from the inherent modularity of distributed energy resources (DERs), such as solar and storage, as well as the persistent growth in end-use loads. This evolution presents significant challenges to system operators who typically lack the tools and processes for managing a complex, distributed power system. To fill this gap, we introduce and implement a Microgrid Operating System (OS), a software platform for monitoring, modeling, and optimizing microgrids and distribution systems. The Microgrid OS is a central layer that links DER hardware, such as batteries, solar, and flexible loads, to energy applications like cost minimization, emissions reduction, and wholesale market participation. The core functions it provides are data acquisition and processing, system modeling and learning, and optimization and control. We present key modules of the Microgrid OS in the context of several implementation projects in microgrids, commercial buildings, and distribution networks.
", "doi": "10.7907/scmm-p028", "publication_date": "2023", "thesis_type": "phd", "thesis_year": "2023" }, { "id": "thesis:14637", "collection": "thesis", "collection_id": "14637", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:05272022-171138586", "primary_object_url": { "basename": "deposit_thesis.pdf", "content": "final", "filesize": 26224719, "license": "other", "mime_type": "application/pdf", "url": "/14637/1/deposit_thesis.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Autonomous Temporal Understanding and State Estimation during Robot-Assisted Surgery", "author": [ { "family_name": "Qin", "given_name": "Yidan", "orcid": "0000-0002-7766-1021", "clpid": "Qin-Yidan" } ], "thesis_advisor": [ { "family_name": "Burdick", "given_name": "Joel W.", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" } ], "thesis_committee": [ { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Tai", "given_name": "Yu-Chong", "orcid": "0000-0001-8529-106X", "clpid": "Tai-Yu-Chong" }, { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Robot-Assisted Surgery (RAS) has become increasingly important in modern surgical practice for its many benefits and advantages for both the patient and the healthcare professionals, as compared to traditional open surgeries and minimally invasive surgeries such as laparoscopy. Artificial intelligence applications during RAS and post-operative analysis can provide various surgeon-assisting functionalities and could potentially achieve a better surgery outcome. These applications, ranging from providing surgeons with advisory information during RAS and post-operative analysis to virtual fixture and supervised autonomous surgical tasks, share a necessary prerequisite of a comprehensive understanding of the current surgical scene. This understanding should include the knowledge of the current surgical task being performed, the surgeon's actions and gestures, the state of the patient, etc. Currently, there is yet to be a unified effort to achieve the autonomous temporal understanding and perception of an RAS at the high accuracy and efficiency required in the highly safety-critical field of medicine.
\r\n\r\nThis thesis develops novel modeling methodologies and deep learning-based models for the autonomous perception and temporal segmentation of the current surgical scene during an RAS. An RAS procedure is modeled as a hierarchical system consisting of discrete surgical states at multiple levels of temporal granularity. These surgical states take the form of surgical tasks, operational steps, fine-grained surgical actions, etc. A broad range of computational experiments were performed to develop methods that achieve an accurate, robust, and efficient estimation of these surgical states. Multiple novel deep learning-based models for feature extraction, noise elimination, and efficient training were proposed and tested. This thesis also shows the significant benefits of incorporating multiple types of data streams recorded by the surgical robotic system to a more accurate surgical state estimation effort.
\r\n\r\nTwo new RAS datasets that contains real-world RAS procedures and diverse experimental settings were collected and annotated--filling a gap in the data sets available for the development and testing of of robust surgical state estimation models. The performance and robustness of models in this thesis work were showcased with these highly complex and dynamic real-world RAS datasets and compared against state-of-the-art methods. A significant model performance improvement was observed in both surgical state estimation accuracy and efficiency. The modeling methodologies and deep learning-based models developed in this work have diverse potential applications to the development of a next-generation surgical robotic systems.
", "doi": "10.7907/n58k-tr61", "publication_date": "2022-06-10", "thesis_type": "phd", "thesis_year": "2022" }, { "id": "thesis:14394", "collection": "thesis", "collection_id": "14394", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:10122021-213903517", "type": "thesis", "title": "Koopman-based Learning and Control of Agile Robotic Systems", "author": [ { "family_name": "Folkestad", "given_name": "Carl A. A.", "orcid": "0000-0002-3436-8247", "clpid": "Folkestad-Carl-A-A" } ], "thesis_advisor": [ { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" } ], "thesis_committee": [ { "family_name": "Ames", "given_name": "Aaron D.", "orcid": "0000-0003-0848-3177", "clpid": "Ames-A-D" }, { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" }, { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Learning methods to enable high performance control systems have recently shown promising results in selected environments and applications. These advances promote the next generation of autonomous robots capable of significantly improving efficiency, cost, and safety in their respective domains. Importantly, these systems are safety-critical and operate in proximity to humans in diverse and uncertain environments. As a result, operational failures may cause significant material and societal losses. Additionally, robot learning and control are further complicated by requiring fast controller update rates and operational constraint satisfaction.
\r\n\r\nTo address these challenges, this thesis presents multiple methods based on Koopman operator theory. The first approach develops algorithms to learn lifted-dimensional models of nonlinear systems and leverages the models in model predictive control (MPC) design. Koopman-based methods typically employ hand-crafted observable functions to \"lift\" the state variables to the higher dimensional space. For most systems, this leads to poor prediction performance and inefficient use of data and computational resources. Instead, I present methods that generate observable functions from data, both based on underlying theory and by incorporating the observable functions and model structure in a neural network model. This allows lower dimensional models, important for real-time control, and enables the nonlinearities of control-affine dynamics to be captured, crucial to describing many robotic systems. I use quadrotor drones to experimentally demonstrate that the learned models combined with MPC can achieve close to optimal behavior while respecting important operational constraints.
\r\n \r\nThe last part of the thesis is concerned with endowing systems with an arbitrary nominal control policy with safety guarantees. Control barrier functions (CBFs) are a powerful tool to achieve this, yet they rely on the computation of control invariant sets, which is notoriously difficult. To avoid this, a backup strategy can be used to implicitly define a control invariant set. However, this requires forward integration of the system dynamics under a backup controller, which is prohibitively expensive for realistic systems. I present a method that replaces the expensive integration using learned Koopman operators of the closed-loop dynamics. As a result, the online computation time required to evaluate the controller is drastically reduced, enabling real-time use. I also derive an error bound on the unmodeled dynamics in order to robustify the CBF controller and demonstrate the method on multi-agent collision avoidance for wheeled robots and quadrotors.
", "doi": "10.7907/2t6d-j206", "publication_date": "2022", "thesis_type": "phd", "thesis_year": "2022" }, { "id": "thesis:14947", "collection": "thesis", "collection_id": "14947", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06062022-021936716", "type": "thesis", "title": "Regret-Optimal Control", "author": [ { "family_name": "Goel", "given_name": "Gautam", "orcid": "000-0002-7054-7218", "clpid": "Goel-Gautam" } ], "thesis_advisor": [ { "family_name": "Hassibi", "given_name": "Babak", "orcid": "0000-0002-1375-5838", "clpid": "Hassibi-B" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Wierman", "given_name": "Adam C.", "clpid": "Wierman-A-C" }, { "family_name": "Hazan", "given_name": "Elad", "orcid": "0000-0002-1566-3216", "clpid": "Hazan-Elad" }, { "family_name": "Hassibi", "given_name": "Babak", "orcid": "0000-0002-1375-5838", "clpid": "Hassibi-B" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Optimal controllers are usually designed to minimize cost under the assumption that the disturbance they encounter is drawn from some specific class. For example, in H\u2082 control the disturbance is assumed to be generated by a stochastic process and the controller is designed to minimize its expected cost, while in H\u221e control the disturbance is assumed to be generated adversarially and the controller is designed to minimize its worst-case cost. This approach suffers from an obvious drawback: a controller which encounters a disturbance which falls outside of the class it was designed to handle may perform poorly. This observation naturally motivates the design of adaptive controllers which dynamically adjust their control strategy as they causally observe the disturbance instead of blindly following a prescribed strategy.
\r\n\r\nInspired by online learning, we propose data-dependent regret as a criterion for controller design. In the regret-optimal control paradigm, causal controllers are designed to minimize regret against a hypothetical optimal noncausal controller, which selects the cost-minimizing sequence of control actions given noncausal access to the disturbance sequence. Controllers with low regret retain a performance guarantee irrespective of how the disturbance is generated; it is this universality which makes our approach an attractive alternative to traditional H\u2082 and H\u221e control. The regret of the causal controller is bounded by some measure of the complexity of the disturbance sequence; we consider several different complexity measures, including the energy of the disturbance sequence, which measures the size of the disturbance, and the pathlength of the disturbance, which measures its variation over time. We also consider the alternative metric of competitive ratio, which is the worst-case ratio between the cost incurred by the causal controller and the cost incurred by the optimal noncausal controller. This metric can also be viewed as a special case of data-dependent regret, where the complexity measure is simply the offline optimal cost. For each of these complexity measures, we derive a corresponding control algorithm with optimal data-dependent regret. The key technique we introduce is an operator-theoretic reduction from regret-optimal control to H\u221e control; each of the regret-optimal controllers we obtain can be interpreted as an H\u221e controller in a synthetic system of larger dimension. We also extend regret-optimal control to the more challenging measurement-feedback setting, where the online controller must choose control actions without directly observing the disturbance sequence, using only noisy linear measurements of the state.
\r\n\r\nWe show that the competitive controller can be arbitrarily well-approximated by the class of disturbance-action-controller (DAC) policies. The convexity of this class of policies makes it amenable to online optimization via a reduction to online convex optimization with memory, and this class has hence attracted much recent attention in online learning. Using our approximation result, we show how to obtain algorithms which achieve the \"best-of-both-worlds\": sublinear policy regret against DAC policies and approximate competitive ratio. These performance guarantees can even be extended to the \"adaptive control\" setting, where the controller does not know the system dynamics ahead of time and must perform online system identification.
\r\n\r\nWe present numerical experiments in a linear dynamical system which demonstrate how the performance of regret-optimal controllers varies as a function of the complexity of the disturbance. We extend regret-optimal control to nonlinear dynamical systems using model-predictive control (MPC) and present experiments which suggest that regret-optimal control is a promising approach to adapting to model error in nonlinear control.
", "doi": "10.7907/8t6d-4t53", "publication_date": "2022", "thesis_type": "phd", "thesis_year": "2022" }, { "id": "thesis:14488", "collection": "thesis", "collection_id": "14488", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:02022022-183557614", "primary_object_url": { "basename": "2022-02_Michael_Maser_phd_thesis.pdf", "content": "final", "filesize": 69533936, "license": "other", "mime_type": "application/pdf", "url": "/14488/1/2022-02_Michael_Maser_phd_thesis.pdf", "version": "v5.0.0" }, "type": "thesis", "title": "Machine Learning Methods Inspired by Challenges in Total Synthesis", "author": [ { "family_name": "Maser", "given_name": "Michael Robert", "orcid": "0000-0001-7895-7804", "clpid": "Maser-Michael-Robert" } ], "thesis_advisor": [ { "family_name": "Reisman", "given_name": "Sarah E.", "orcid": "0000-0001-8244-9300", "clpid": "Reisman-S-E" } ], "thesis_committee": [ { "family_name": "Stoltz", "given_name": "Brian M.", "orcid": "0000-0001-9837-1528", "clpid": "Stoltz-B-M" }, { "family_name": "Reisman", "given_name": "Sarah E.", "orcid": "0000-0001-8244-9300", "clpid": "Reisman-S-E" }, { "family_name": "Fu", "given_name": "Gregory C.", "orcid": "0000-0002-0927-680X", "clpid": "Fu-G-C" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Listgarten", "given_name": "Jennifer", "orcid": "0000-0002-6600-1431", "clpid": "Listgarten-Jennifer" } ], "local_group": [ { "literal": "div_chem" } ], "abstract": "Synthetic organic chemists face a dearth of challenges in the efficient construction of functional molecules, particularly bioactive compounds. Predictive approaches offer reductions to research timelines and resource costs and allow chemists to devote their expertise where it is most valuable. Promising machine learning (ML) methods have evolved for uncovering patterns in chemical data that are beyond the grasp of expert humans, but a number of grand challenges in molecular ML remain. First, the learning of chemical structure representations rooted in physical first principles has yet to be robustly demonstrated. Second, the practical task of predicting successful \"over-the-arrow\" reaction conditions remains elusive. Finally, the demonstration of such solutions in the context of complex synthesis has yet to be realized.
\r\n\r\nHerein, approaches to these grand challenges are developed and described. Inspiration is derived from the successful synthesis of the anticancer marine natural product ritterazine B. Reaction condition prediction is approached first, where a novel graph neural network architecture is developed under a multilabel classification framework. The resulting model is successfully demonstrated on datasets of four high-value reaction types in modern synthesis. Next, 3D-to-1D representation learning is approached by development of a volumetric neural architecture based on inception networks. Such voxel models are demonstrated for the prediction of expensive quantum mechanical properties from space-filled data alone.
\r\n\r\nThe merging of these approaches for reaction condition optimization and utility in complex settings is discussed and forecasted for future works, which are currently underway.
", "doi": "10.7907/ffjk-g059", "publication_date": "2022", "thesis_type": "phd", "thesis_year": "2022" }, { "id": "thesis:14357", "collection": "thesis", "collection_id": "14357", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:09092021-024951318", "primary_object_url": { "basename": "Aceves_Aiden_2021_Thesis_Revised_v2.pdf", "content": "final", "filesize": 9915385, "license": "other", "mime_type": "application/pdf", "url": "/14357/2/Aceves_Aiden_2021_Thesis_Revised_v2.pdf", "version": "v9.0.0" }, "type": "thesis", "title": "Machine Learning and Modeling Methods for Protein Engineering", "author": [ { "family_name": "Aceves", "given_name": "Aiden Joseph", "orcid": "0000-0001-6583-5373", "clpid": "Aceves-Aiden-Joseph" } ], "thesis_advisor": [ { "family_name": "Mayo", "given_name": "Stephen L.", "orcid": "0000-0002-9785-5018", "clpid": "Mayo-S-L" } ], "thesis_committee": [ { "family_name": "Bjorkman", "given_name": "Pamela J.", "orcid": "0000-0002-2277-3990", "clpid": "Bjorkman-P-J" }, { "family_name": "Miller", "given_name": "Thomas F.", "orcid": "0000-0002-1882-5380", "clpid": "Miller-T-F" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Van Valen", "given_name": "David A.", "orcid": "0000-0001-7534-7621", "clpid": "Van-Valen-D" }, { "family_name": "Mayo", "given_name": "Stephen L.", "orcid": "0000-0002-9785-5018", "clpid": "Mayo-S-L" } ], "local_group": [ { "literal": "div_bbe" } ], "abstract": "Computation has been an integral part of structural biology, ever since the first protein macromolecular structure was solved via Fourier Synthesis on the EDSAC Mark I electronic computer in 1958 (Kendrew et al., 1958). Throughout my time at Caltech, I have endeavored to develop new methods to apply machine learning and molecular modeling to the study of biological macromolecules. These efforts have taken two distinct tracks, but are unified by a focus on studying proteins on a structural level.
\r\n\r\nThrough the application of molecular dynamics and modeling, I have studied insulin from several angles, including the incorporation of non-canonical amino acids, and how these modifications might be responsible for the modification of critical properties such as hexamer dissociation and fibrillation formation. Additionally, I have probed how insulin behaves at the interface of water and silica, a property which is critical for the effective dissemination and administration of this therapeutic molecule. I have helped to develop a novel computationally guided workflow for integrating drug conjugates into antibody CDRs. This technique yields molecules which exhibit synergistic binding and an enhanced ability for selective binding.
\r\n\r\nThe second major thrust of my research has focused on applying machine learning to protein engineering problems, particularly developing tools for working with structural data, and for making efficient re-use of data which has already been laboriously collected by other groups. The basic data parsing and processing tools which were created and refined over the course of my time at Caltech has enabled many other projects, both of my own and of collaborators. Studies into the use of generative networks for protein-protein docking have been conducted which lend useful insights for network architecture, the inclusion of intermediate learning objectives, and overcoming sparsity. The technique introduced in our ICLR 2021 paper demonstrates a regularization method which enables data from past protein engineering campaigns to be leveraged to learn policies which optimally select molecules to synthesize in unrelated engineering efforts, to potentially save a significant amount of time and money for future projects.
\r\n\r\nReference
\r\n\r\nKendrew, J. C.; Bodo, G.; Dintzis, H. M.; Parrish, R. G.; Wyckoff, H.; Phillips, D. C. A. \"Three-Dimensional Model of the Myoglobin Molecule Obtained by X-Ray Analysis\". Nature 1958, 181 (4610), 662\u2013666.
", "doi": "10.7907/5j34-0d55", "publication_date": "2022", "thesis_type": "phd", "thesis_year": "2022" }, { "id": "thesis:14939", "collection": "thesis", "collection_id": "14939", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06032022-035416994", "type": "thesis", "title": "Cascading Failures in Power Systems: Modeling, Characterization, and Mitigation", "author": [ { "family_name": "Liang", "given_name": "Chen", "orcid": "0000-0002-0015-7206", "clpid": "Liang-Chen" } ], "thesis_advisor": [ { "family_name": "Low", "given_name": "Steven H.", "orcid": "0000-0001-6476-3048", "clpid": "Low-S-H" }, { "family_name": "Wierman", "given_name": "Adam C.", "clpid": "Wierman-A-C" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Low", "given_name": "Steven H.", "orcid": "0000-0001-6476-3048", "clpid": "Low-S-H" }, { "family_name": "Wierman", "given_name": "Adam C.", "clpid": "Wierman-A-C" }, { "family_name": "Chandrasekaran", "given_name": "Venkat", "clpid": "Chandrasekaran-V" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Reliability is a critical goal for power systems. Due to the connectivity of power grids, an initial failure may trigger a cascade of failures and eventually lead to a large-scale blackout, causing significant economic and social impacts. Cascading failure analysis thus draws wide attention from power system practitioners and researchers. A well-known observation is that cascading failures in power systems propagate non-locally because of the complex mechanism of power grids. Such non-local propagation makes it particularly challenging to model, analyze and control the failure process. In this thesis, we tackle these challenges by establishing a mathematical theory to model and characterize failure patterns, discover structural properties of failure propagation, and design novel techniques for failure mitigation.
\r\n\r\nFirst, we propose a failure propagation model considering both fast-timescale system frequency dynamics and the slow-timescale line tripping process. This model provides mathematical justifications to the widely used static DC model and can be generalized to capture a variety of failure propagation patterns induced by different control mechanisms of the power grid. More importantly, this model provides flexibility to design real-time control algorithms for failure mitigation and localization.
\r\n\r\nSecond, we provide a complete characterization of line failures under the static DC model. Our results unveil a deep connection between the power redistribution patterns and the network block decomposition. More specifically, we show that a non-cut line failure in a block will only impact the branch power flows on the transmission lines within the block. In contrast, a cut set line failure will propagate globally depending on both the power balancing rules and the network topological structure. Further, we discuss three types of interface networks to connect the sub-grids, all achieving better failure localization performance.
\r\n\r\nThird, we study corrective control algorithms for failure mitigation. We integrate a distributed frequency control strategy with the network block decomposition to provide provable failure mitigation and localization guarantees on line failures. This strategy operates on the frequency control timescale and supplements existing corrective mechanisms, improving grid reliability and operational efficiency. We further explore the failure mitigation approach with direct post-contingency injection adjustments. Specifically, we propose an optimization-based control method with strong structural properties, which is highly desirable in large-scale power networks.
", "doi": "10.7907/8817-xy25", "publication_date": "2022", "thesis_type": "phd", "thesis_year": "2022" }, { "id": "thesis:14537", "collection": "thesis", "collection_id": "14537", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:04012022-011925533", "primary_object_url": { "basename": "Dissertation_Logan_Cross.pdf", "content": "final", "filesize": 60965298, "license": "other", "mime_type": "application/pdf", "url": "/14537/1/Dissertation_Logan_Cross.pdf", "version": "v5.0.0" }, "type": "thesis", "title": "The Neural Mechanisms of Value Construction", "author": [ { "family_name": "Cross", "given_name": "Logan Matthew", "orcid": "0000-0002-5248-9499", "clpid": "Cross-Logan-Matthew" } ], "thesis_advisor": [ { "family_name": "O'Doherty", "given_name": "John P.", "orcid": "0000-0003-0016-3531", "clpid": "O'Doherty-J-P" } ], "thesis_committee": [ { "family_name": "Adolphs", "given_name": "Ralph", "orcid": "0000-0002-8053-9692", "clpid": "Adolphs-R" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Mobbs", "given_name": "Dean", "orcid": "0000-0003-1175-3772", "clpid": "Mobbs-Dean" }, { "family_name": "O'Doherty", "given_name": "John P.", "orcid": "0000-0003-0016-3531", "clpid": "O'Doherty-J-P" } ], "local_group": [ { "literal": "div_bbe" } ], "abstract": "Research in decision neuroscience has characterized how the brain makes decisions by assessing the expected utility of each option in an abstract value space that affords the ability to compare dissimilar options. Experiments at multiple levels of analysis in multiple species have localized the ventromedial prefrontal cortex (vmPFC) and nearby orbitofrontal cortex (OFC) as the main nexus where this abstract value space is represented. However, much less is known about how this value code is constructed by the brain in the first place. By using a combination of behavioral modeling and cutting edge tools to analyze functional magnetic resonance imaging (fMRI) data, the work of this thesis proposes that the brain decomposes stimuli into their constituent attributes and integrates across them to construct value. These stimulus features embody appetitive or aversive properties that are either learned from experience or evaluated online by comparing them to previously experienced stimuli with similar features. Stimulus features are processed by cortical areas specialized for the perception of a particular stimulus type and then integrated into a value signal in vmPFC/OFC.
\r\n\r\nThe project presented in Chapter 2 examines how food items are evaluated by their constituent attributes, namely their nutrient makeup. A linear attribute integration model succinctly captures how subjective values can be computed from a weighted combination of the constituent nutritive attributes of the food. Multivariate analysis methods revealed that these nutrient attributes are represented in the lateral OFC, while food value is encoded both in medial and lateral OFC. Connectivity between lateral and medial OFC allows this nutrient attribute information to be integrated into a value representation in medial OFC.
\r\n\r\nIn Chapter 3, I show that this value construction process can operate over higher-level abstractions when the context requires bundles of items to be valued, rather than isolated items. When valuing bundles of items, the constituent items themselves become the features, and their values are integrated with a subadditive function to construct the value of the bundle. Multiple subregions of PFC including but not limited to vmPFC compute the value of a bundle with the same value code used to evaluate individual items, suggesting that these general value regions contextually adapt within this hierarchy. When valuing bundles and single items in interleaved trials, the value code rapidly switches between levels in this hierarchy by normalizing to the distribution of values in the current context rather than representing all options on an absolute scale.
\r\n\r\nAlthough the attribute integration model of value construction characterizes human behavior on simple decision-making tasks, it is unclear how it can scale up to environments of real-world complexity. Taking inspiration from modern advances in artificial intelligence, and deep reinforcement learning in particular, in Chapter 4 I outline how connectionist models generalize the attribute integration model to naturalistic tasks by decomposing sensory input into a high dimensional set of nonlinear features that are encoded with hierarchical and distributed processing. Participants freely played Atari video games during fMRI scanning, and a deep reinforcement learning algorithm trained on the games was used as an end-to-end model for how humans evaluate actions in these high-dimensional tasks. The features represented in the intermediate layers of the artificial neural network were found to also be encoded in a distributed fashion throughout the cortex, specifically in the dorsal visual stream and posterior parietal cortex. These features emerge from nonlinear transformations of the sensory input that connect perception to action and reward. In contrast to the stimulus attributes used to evaluate the stimuli presented in the preceding chapters, these features become highly complex and inscrutable as they are driven by the statistical properties of high-dimensional data. However, they do not solely reflect a set of features that can be identified by applying common dimensionality reduction techniques to the input, as task-irrelevant sensory features are stripped away and task-relevant high-level features are magnified.
", "doi": "10.7907/x5kk-8h27", "publication_date": "2022", "thesis_type": "phd", "thesis_year": "2022" }, { "id": "thesis:14496", "collection": "thesis", "collection_id": "14496", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:02132022-064810187", "primary_object_url": { "basename": "David_Brown_Thesis_V4.pdf", "content": "final", "filesize": 10109719, "license": "other", "mime_type": "application/pdf", "url": "/14496/1/David_Brown_Thesis_V4.pdf", "version": "v5.0.0" }, "type": "thesis", "title": "Principles of Massively Parallel Sequencing for Engineering and Characterizing Gene Delivery", "author": [ { "family_name": "Brown", "given_name": "David", "orcid": "0000-0002-9757-1744", "clpid": "Brown-David" } ], "thesis_advisor": [ { "family_name": "Gradinaru", "given_name": "Viviana", "orcid": "0000-0001-5868-348X", "clpid": "Gradinaru-V" }, { "family_name": "Thomson", "given_name": "Matthew", "orcid": "0000-0003-1021-1234", "clpid": "Thomson-M-W" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Arnold", "given_name": "Frances Hamilton", "orcid": "0000-0002-4027-364X", "clpid": "Arnold-F-H" }, { "family_name": "Gradinaru", "given_name": "Viviana", "orcid": "0000-0001-5868-348X", "clpid": "Gradinaru-V" }, { "family_name": "Thomson", "given_name": "Matthew", "orcid": "0000-0003-1021-1234", "clpid": "Thomson-M-W" } ], "local_group": [ { "literal": "div_bbe" } ], "abstract": "The advent of massively parallel sequencing and synthesis technologies have ushered in a new paradigm of biology, where high throughput screening of billions of nucleid acid molecules and production of libraries of millions of genetic mutants are now routine in labs and clinics. During my Ph.D., I worked to develop data analysis and experimental methods that take advantage of the scale of this data, while making the minimal assumptions necessary for deriving value from their application. My Ph.D. work began with the development of software and principles for analyzing deep mutational scanning data of libraries of engineered AAV capsids. By looking at not only the top variant in a round of directed evolution, but instead a broad distribution of the variants and their phenotypes, we were able to identify AAV variants with enhanced ability to transduce specific cells in the brain after intravenous injection. I then shifted to better understand the phenotypic profile of these engineered variants. To that end, I turned to single-cell RNA sequencing to seek to identify, with high resolution, the delivery profile of these variants in all cell types present in the cortex of a mouse brain. I began by developing infrastructure and tools for dealing with the data analysis demands of these experiments. Then, by delivering an engineered variant to the animal, I was able to use the single-cell RNA sequencing profile, coupled with a sequencing readout of the delivered genetic cargo present in each cell type, to define the variant\u2019s tropism across the full spectrum of cell types in a single step. To increase the throughput of this experimental paradigm, I then worked to develop a multiplexing strategy for delivering up to 7 engineered variants in a single animal, and obtain the same high resolution readout for each variant in a single experiment. Finally, to take a step towards translation to human diagnostics, I leveraged the tools I built for scaling single-cell RNA sequencing studies and worked to develop a protocol for obtaining single-cell immune profiles of low volumes of self-collected blood. This study enabled repeat sampling in a short period of time, and revealed an incredible richness in individual variability and time-of-day dependence of human immune gene expression. Together, my Ph.D. work provides strategies for employing massively parallel sequencing and synthesis for new biological applications, and builds towards a future paradigm where personalized, high-resolution sequencing might be coupled with modular, customized gene therapy delivery.
", "doi": "10.7907/yqjm-6609", "publication_date": "2022", "thesis_type": "phd", "thesis_year": "2022" }, { "id": "thesis:14489", "collection": "thesis", "collection_id": "14489", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:02062022-203334930", "primary_object_url": { "basename": "Ben_Stevens_Thesis (4).pdf", "content": "final", "filesize": 5499929, "license": "other", "mime_type": "application/pdf", "url": "/14489/1/Ben_Stevens_Thesis (4).pdf", "version": "v5.0.0" }, "type": "thesis", "title": "Applications of Machine Learning to Finite Volume Methods", "author": [ { "family_name": "Stevens", "given_name": "Benjamin Carter", "orcid": "0000-0002-3410-5922", "clpid": "Stevens-Benjamin-Carter" } ], "thesis_advisor": [ { "family_name": "Colonius", "given_name": "Tim", "orcid": "0000-0003-0326-3909", "clpid": "Colonius-T" } ], "thesis_committee": [ { "family_name": "Blanquart", "given_name": "Guillaume", "orcid": "0000-0002-5074-9728", "clpid": "Blanquart-G" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Anandkumar", "given_name": "Anima", "orcid": "0000-0002-6974-6797", "clpid": "Anandkumar-A" }, { "family_name": "Colonius", "given_name": "Tim", "orcid": "0000-0003-0326-3909", "clpid": "Colonius-T" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "The finite volume method (FVM) has been one of the primary tools of computational fluid dynamics (CFD) for many decades. This method allows for the approximate solution of a partial differential equation (PDE) to be determined by breaking up a problem with no analytical solution into smaller pieces that can be solved together to get a physically realistic simulation. These algorithms can even be used for PDEs with discontinuous solutions, though they must be carefully designed for those situations because they cannot assume any level of smoothness in the solution. An FVM that has been designed for PDEs with discontinuous solutions is referred to as a shock-capturing method. For most of their history, FVM algorithms have been developed using rigorous mathematical arguments to formally maximize the order of convergence of the solution as the grid is refined. However, these arguments depend on the solution to the PDE being smooth, and therefore do not apply to shock-capturing methods. Instead, shock-capturing methods have traditionally been designed using human intuition to create algorithms that then perform well empirically. In this thesis, we instead follow a data-driven approach to train neural networks to use for enhanced FVM methods.
\r\n\r\nBy including a neural network in our FVM, we can use empirical data to optimize the algorithm. We can also utilize ideas from traditional FVM algorithms to create hybrid methods that have tunable parameters and maintain convergence guarantees present in FVMs that have been designed by hand. We explore these hybrid methods in a variety of settings. First, we create a general-purpose shock-capturing method WENO-NN by hybridizing the popular shock-capturing method WENO-JS with a neural network. Additionally, we develop a network architecture, called FiniteNet, that can be used to learn a coarse-graining model associated with a specific PDE and embed it into an FVM scheme. Finally, we also explore the idea of using transfer learning to further improve the WENO-NN for specific problems and name the resulting algorithm WENO-TL. We demonstrate experimentally that this hybrid approach results in methods that can offer similar error levels as traditional FVMs at less computational cost. Although the neural network increases the computational cost of one evaluation of our hybrid FVM, these methods also allow the simulation to be carried out on a coarser grid, leading to a net reduction in both simulation time and memory usage.
", "doi": "10.7907/41qn-7n22", "publication_date": "2022", "thesis_type": "phd", "thesis_year": "2022" }, { "id": "thesis:14449", "collection": "thesis", "collection_id": "14449", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:12152021-005204320", "type": "thesis", "title": "Biological Intelligence: from Behavior to Learning Theory", "author": [ { "family_name": "Zhang", "given_name": "Tony (Haoyu)", "orcid": "0000-0002-5198-499X", "clpid": "Zhang-Tony-Haoyu" } ], "thesis_advisor": [ { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" }, { "family_name": "Meister", "given_name": "Markus", "orcid": "0000-0003-2136-6506", "clpid": "Meister-M" }, { "family_name": "Bouman", "given_name": "Katherine L.", "orcid": "0000-0003-0077-4367", "clpid": "Bouman-K-L" } ], "local_group": [ { "literal": "div_bbe" } ], "abstract": "Knowing how to learn, think, and act is not just a hallmark of intelligence, but a necessity of survival for many organisms. Behavior, the complete set of actions of species, allows us to glimpse into the minds of humans and animals, and by extension, intelligence itself. Biological intelligence is characterized by fast adaptation to changes and challenges, which is what allows species to survive in natural environments from starvation and predation. To study learning in a controlled setting, we can observe the behavior evoked through decision-making tasks that make it possible to quantify and analyze learning. By modeling the extracted behavioral features, we could start to understand the possible underlying mechanisms by proposing neural theory models, and look for those signals in the brain. Understanding the neural mechanisms of learning also strengthens the basis for building intelligent machines that are flexible and adaptive to the nonstationary world we live in. In this thesis, I present works in (1) automating behavioral setups and modeling suboptimal behavior in a traditional decision-making task, (2) using an ethological navigation task to characterize fast-sequence learning, and (3) how neural theory can explain some core behavioral phenomena in (2), and be used to solve a central problem in graph search.
", "doi": "10.7907/8z4q-8g15", "publication_date": "2022", "thesis_type": "phd", "thesis_year": "2022" }, { "id": "thesis:14436", "collection": "thesis", "collection_id": "14436", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:11302021-224628633", "primary_object_url": { "basename": "zhan_eric_2022_thesis.pdf", "content": "final", "filesize": 8876717, "license": "other", "mime_type": "application/pdf", "url": "/14436/1/zhan_eric_2022_thesis.pdf", "version": "v5.0.0" }, "type": "thesis", "title": "New Algorithms for Programmatic Deep Learning with Applications to Behavior Modeling", "author": [ { "family_name": "Zhan", "given_name": "Eric", "orcid": "0000-0003-0521-5329", "clpid": "Zhan-Eric" } ], "thesis_advisor": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "thesis_committee": [ { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" }, { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" }, { "family_name": "Chaudhuri", "given_name": "Swarat", "orcid": "0000-0002-6859-1391", "clpid": "Chaudhuri-Swarat" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Raw behavioral data is becoming increasingly more abundant and more easily obtainable in spatiotemporal domains such as sports, video games, navigation & driving, motion capture, and animal science. How can we best use this data to advance their respective domains forward? For instance, researchers for self-driving vehicles would like to identify the key features of the environment state that impact decision-making the most; game developers would like to populate their games with characters that have unique and diverse behaviors to create a more immersive gaming experience; and behavioral neuroscientists would like to uncover the underlying mechanisms that drive learning in animals. Machine learning, the science of developing models and algorithms to identify and leverage patterns in data, is well-equipped to aid in these endeavors. But how do we integrate machine learning with these spatiotemporal domains in a principled way? In this dissertation, we develop and introduce new algorithms in programmatic deep learning that tackle some of the new challenges encountered in behavior modeling.
\r\n\r\nOur work in programmatic deep learning comprises two main themes: in the first, we show how to use expert-written programs as sources of weak labels in domains where manually-annotated expert labels are scarce; in the second, we explore programs as a flexible function class with human-interpretable structure and show how to learn them via neurosymbolic program learning. Augmenting deep learning with programmatic structure allows domain experts to easily incorporate domain knowledge into machine learning models; we show that this results in significant improvements in many behavior modeling applications like imitation learning, controllable generation, counterfactual analysis, and unsupervised clustering.
", "doi": "10.7907/5n5q-x203", "publication_date": "2022", "thesis_type": "phd", "thesis_year": "2022" }, { "id": "thesis:14045", "collection": "thesis", "collection_id": "14045", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:01042021-220032574", "type": "thesis", "title": "Data-Driven Protein Engineering", "author": [ { "family_name": "Wu", "given_name": "Zachary", "orcid": "0000-0003-2429-9812", "clpid": "Wu-Zachary" } ], "thesis_advisor": [ { "family_name": "Arnold", "given_name": "Frances Hamilton", "orcid": "0000-0002-4027-364X", "clpid": "Arnold-F-H" } ], "thesis_committee": [ { "family_name": "Tirrell", "given_name": "David A.", "orcid": "0000-0003-3175-4596", "clpid": "Tirrell-D-A" }, { "family_name": "Wang", "given_name": "Zhen-Gang", "orcid": "0000-0002-3361-6114", "clpid": "Wang-Zhen-Gang" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Arnold", "given_name": "Frances Hamilton", "orcid": "0000-0002-4027-364X", "clpid": "Arnold-F-H" } ], "local_group": [ { "literal": "div_chem" } ], "abstract": "Directed evolution has enabled the adaptation of natural protein sequences for an endless variety of human applications. Given a starting point - a sequence with measurable activity - directed evolution is able to improve protein sequences by iteratively accumulating beneficial mutations. However, directed evolution requires investing large experimental effort, which continues to be the major bottleneck in efficient protein optimization. To this end, we describe a framework for incorporating machine learning in the directed evolution process to maximize the utility of generated experimental data in Chapter 2. In Chapter 3, we then show that this framework outperforms traditional directed evolution methods on an empirical fitness landscape. However, directed evolution is fundamentally limited by its need for a starting point, or a sequence with measurable activity. To tackle this issue, we test the ability of nascent deep learning techniques for generating short, functional amino acid sequences in Chapter 4. Encouraged by this success, we attempted to generate full length enzymatic sequences for desired substrates without success. However, we were able to apply this deep learning approach to model other aspects of enzymatic protein sequences in Chapter 5. Finally, the field of data-driven protein sequence generation is enjoying a recent surge in interest, and we provide an updated review of protein engineering with machine learning, focusing on recent work in deep generative modeling in Chapter 1.
", "doi": "10.7907/nx3c-qb44", "publication_date": "2021", "thesis_type": "phd", "thesis_year": "2021" }, { "id": "thesis:14234", "collection": "thesis", "collection_id": "14234", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06022021-223508132", "primary_object_url": { "basename": "Thesis__Learning_to_Optimize (8).pdf", "content": "final", "filesize": 3672012, "license": "other", "mime_type": "application/pdf", "url": "/14234/1/Thesis__Learning_to_Optimize (8).pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Learning to Optimize: from Theory to Practice", "author": [ { "family_name": "Song", "given_name": "Jialin", "orcid": "0000-0001-5633-9909", "clpid": "Song-Jialin" } ], "thesis_advisor": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "thesis_committee": [ { "family_name": "Wierman", "given_name": "Adam C.", "clpid": "Wierman-A-C" }, { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Dilkina", "given_name": "Bistra", "orcid": "0000-0002-6784-473X", "clpid": "Dilkina-Bistra" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Optimization is at the heart of everyday applications, from finding the fastest route for navigation to designing efficient drugs for diseases. The study of optimization algorithms has focused on developing general approaches that do not adapt to specific problem instances. While they enjoy wide applicability, they forgo the potentially useful information embedded in the structure of an instance. Furthermore, as new optimization problems appear, the algorithm development process relies heavily on domain expertise to identify special properties and design methods to exploit them. Such design philosophy is labor-intensive and difficult to deploy efficiently to a broad range of domain-specific optimization problems, which are becoming ubiquitous in the pursuit of ever more personalized applications.
\r\n \r\nIn this dissertation, we consider different hybrid versions of classical optimization algorithms with data-driven techniques. We aim to equip classical algorithms with the ability to adapt their behaviors on the fly based on specific problem instances. A common theme in our approaches is to train the data-driven components on a pre-collected batch of representative problem instances to optimize some performance metrics, e.g., wall-clock time. Varying the integration details, we present several approaches to learning data-driven optimization modules for combinatorial optimization problems and study the corresponding fundamental research questions on policy learning. We provide multiple practical experimental results to showcase the practicality of our methods which lead to state-of-the-art performance on some classes of problems.
", "doi": "10.7907/7qaw-kd75", "publication_date": "2021", "thesis_type": "phd", "thesis_year": "2021" }, { "id": "thesis:14021", "collection": "thesis", "collection_id": "14021", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:12092020-162149429", "type": "thesis", "title": "Online Learning from Human Feedback with Applications to Exoskeleton Gait Optimization", "author": [ { "family_name": "Novoseller", "given_name": "Ellen Rachel", "orcid": "0000-0001-5263-0598", "clpid": "Novoseller-Ellen-Rachel" } ], "thesis_advisor": [ { "family_name": "Burdick", "given_name": "Joel W.", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "thesis_committee": [ { "family_name": "Ames", "given_name": "Aaron D.", "orcid": "0000-0003-0848-3177", "clpid": "Ames-A-D" }, { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" }, { "family_name": "Sadigh", "given_name": "Dorsa", "orcid": "0000-0002-7802-9183", "clpid": "Sadigh-Dorsa" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Systems that intelligently interact with humans could improve people's lives in numerous ways and in numerous settings, such as households, hospitals, and workplaces. Yet, developing algorithms that reliably and efficiently personalize their interactions with people in real-world environments remains challenging. In particular, one major difficulty lies in adapting to human-in-the-loop feedback, in which an algorithm makes sequential decisions while receiving online feedback from humans; throughout this interaction, the algorithm seeks to optimize its decision-making quality, as measured by the utility of its performance to the human users. Such algorithms must balance between exploration and exploitation: on one hand, the algorithm must select uncertain strategies to fully explore the environment and the interacting human's preferences, while on the other hand, it must exploit the empirically-best-performing strategies to maximize its cumulative performance.
\r\n\r\nLearning from human feedback can be difficult, as people are often unreliable in specifying numerical scores. In contrast, humans can often more accurately provide various types of qualitative feedback, for instance pairwise preferences. Yet, sample efficiency is a significant concern in human-in-the-loop settings, as qualitative feedback is less informative than absolute metrics, and algorithms can typically pose only limited queries to human users. Thus, there is a need to create theoretically-grounded online learning algorithms that efficiently, reliably, and robustly optimize their interactions with humans while learning from online qualitative feedback.
\r\n\r\nThis dissertation makes several contributions to algorithm design for human-in-the-loop learning. Firstly, this work develops the Dueling Posterior Sampling (DPS) algorithmic framework, a model-based, Bayesian approach for online learning in the settings of preference-based reinforcement learning and generalized linear dueling bandits. DPS is developed together with a theoretical regret analysis framework, and yields competitive empirical performance in a range of simulations. Additionally, this thesis presents the CoSpar and LineCoSpar algorithms for sample-efficient, mixed-initiative learning from pairwise preferences and coactive feedback. CoSpar and LineCoSpar are both deployed in human subject experiments with a lower-body exoskeleton to identify optimal, user-preferred exoskeleton walking gaits. This work presents the first demonstration of preference-based learning for optimizing dynamic crutchless exoskeleton walking for user comfort, and makes progress toward customizing exoskeletons and other assistive devices for individual users.
", "doi": "10.7907/gvtx-1586", "publication_date": "2021", "thesis_type": "phd", "thesis_year": "2021" }, { "id": "thesis:14192", "collection": "thesis", "collection_id": "14192", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:05282021-174607976", "primary_object_url": { "basename": "bagherian_dawna_2021.pdf", "content": "final", "filesize": 8166683, "license": "other", "mime_type": "application/pdf", "url": "/14192/3/bagherian_dawna_2021.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Artificial Neural Networks for Nonlinear System Identification of Neuronal Microcircuits", "author": [ { "family_name": "Bagherian", "given_name": "Dawna Paria", "orcid": "0000-0003-4465-552X", "clpid": "Bagherian-Dawna-Paria" } ], "thesis_advisor": [ { "family_name": "Meister", "given_name": "Markus", "orcid": "0000-0003-2136-6506", "clpid": "Meister-M" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Tsao", "given_name": "Doris Y.", "orcid": "0000-0003-1083-1919", "clpid": "Tsao-D-Y" }, { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" }, { "family_name": "Meister", "given_name": "Markus", "orcid": "0000-0003-2136-6506", "clpid": "Meister-M" } ], "local_group": [ { "literal": "div_bbe" } ], "abstract": "This thesis explores the application of artificial neural networks (ANNs) to nonlinear system identification. We use neuronal microcircuits in the retina as a testbed for our technique, which relies upon the marriage of partial anatomical information with large electrophysiological datasets. Rather than a typical application of machine learning, our primary goal is not to predict the output of retinal circuits, but rather to uncover their structure. We begin with a theoretical exploration in a toy problem and provide a proof of unique identifiability under a specific set of conditions. We then perform empirical simulations in a number of different circuit architectures and explore the space of constraints and regularizers to demonstrate that this technique is feasible in a hyperparametric regime that lends itself well to neuroscience datasets. We then apply the technique to mouse retinal datasets and show that we can both recover known biological information as well as discover new hypotheses for biological exploration. We end with an exploration of active stimulus design algorithms to distinguish between circuit hypotheses.
", "doi": "10.7907/rj2p-8g11", "publication_date": "2021", "thesis_type": "phd", "thesis_year": "2021" }, { "id": "thesis:14254", "collection": "thesis", "collection_id": "14254", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06072021-172842077", "primary_object_url": { "basename": "THESIS_21__Yu_Su.pdf", "content": "final", "filesize": 1940103, "license": "other", "mime_type": "application/pdf", "url": "/14254/1/THESIS_21__Yu_Su.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Optimizing Cloud AI Platforms: Resource Allocation and Market Design", "author": [ { "family_name": "Su", "given_name": "Yu", "orcid": "0000-0002-7159-4542", "clpid": "Su-Yu" } ], "thesis_advisor": [ { "family_name": "Wierman", "given_name": "Adam C.", "clpid": "Wierman-A-C" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Low", "given_name": "Steven H.", "orcid": "0000-0001-6476-3048", "clpid": "Low-S-H" }, { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" }, { "family_name": "Wierman", "given_name": "Adam C.", "clpid": "Wierman-A-C" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "The numerous applications of data-driven algorithms and tools across diverse industries have led to tremendous successes in recent years. As the volume of massive data that is created, collected, and consumed continues to grow, there are many new imposed challenges faced by today's cloud AI platforms that support the deployment of machine learning algorithms on a large scale. In this thesis, we tackle the emerging challenges within cloud AI systems and beyond by adopting approaches from the fields of resource allocation and market design.
\r\n\r\nFirst, we propose a new scheduler, Generalized Earliest Time First (GETF), and provide the provable, worst-case approximation guarantees for the goals of minimizing both makespan and total weighted completion time of tasks with precedence constraints on related machines with machine-dependent communication times. These two results address long-standing open problems. Further, we adopt the classic speed scaling function to model power consumption and use mean response time to measure the performance. We propose the concept of pseudo-size to quantify importance of tasks and design a family of two-stage scheduling frameworks based on the approximation of pseudo-size. Assuming a good approximation of pseudo-size, we are able to provide the first provable bound of a linear combination of performance and energy goals under this setting.
\r\n\r\nSecond, we study the design of mechanisms for data acquisition in settings with information leakage and verifiable data. We provide the first characterization of an optimal mechanism for data acquisition if agents are concerned about privacy and their data is correlated with each other. Additionally, the mechanism allows, for the first time, a trade-off between the bias and variance of the estimator. Transitioning from the data market into the energy market, we propose a new pricing scheme, which is applicable to general non-convex costs, and allows using general parametric pricing functions. Optimizing for the quantities and the price parameters simultaneously, and the ability to use general parametric pricing functions allows our scheme to find prices that are typically economically more efficient and less discriminatory than those of the existing schemes while still supporting a competitive equilibrium. In addition, we supplement the proposed method with a computationally efficient polynomial-time approximation algorithm, which can be used to approximate the optimal quantities and prices for general non-convex cost functions.
", "doi": "10.7907/bc2t-er54", "publication_date": "2021", "thesis_type": "phd", "thesis_year": "2021" }, { "id": "thesis:14178", "collection": "thesis", "collection_id": "14178", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:05272021-042158260", "primary_object_url": { "basename": "joseph_marino_Caltech_PhD_Thesis.pdf", "content": "final", "filesize": 17694235, "license": "other", "mime_type": "application/pdf", "url": "/14178/1/joseph_marino_Caltech_PhD_Thesis.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Learned Feedback & Feedforward Perception & Control", "author": [ { "family_name": "Marino", "given_name": "Joseph Louis", "orcid": "0000-0001-6387-8062", "clpid": "Marino-Joseph-Louis" } ], "thesis_advisor": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" }, { "family_name": "O'Doherty", "given_name": "John P.", "orcid": "0000-0003-0016-3531", "clpid": "O'Doherty-J-P" }, { "family_name": "Tsao", "given_name": "Doris Y.", "orcid": "0000-0003-1083-1919", "clpid": "Tsao-D-Y" }, { "family_name": "Rao", "given_name": "Rajesh P. N.", "clpid": "Rao-Rajesh-P-N" } ], "local_group": [ { "literal": "div_bbe" } ], "abstract": "The notions of feedback and feedforward information processing gained prominence under cybernetics, an early movement at the dawn of computer science and theoretical neuroscience. Negative feedback processing corrects errors, whereas feedforward processing makes predictions, thereby preemptively reducing errors. A key insight of cybernetics was that such processes can be applied to both perception, or state estimation, and control, or action selection. The remnants of this insight are found in many modern areas, including predictive coding in neuroscience and deep latent variable models in machine learning. This thesis draws on feedback and feedforward ideas developed within predictive coding, adapting them to improve machine learning techniques for perception (Part II) and control (Part III). Upon establishing these conceptual connections, in Part IV, we traverse this bridge, from machine learning back to neuroscience, arriving at new perspectives on the correspondences between these fields.
", "doi": "10.7907/4mjd-ce53", "publication_date": "2021", "thesis_type": "phd", "thesis_year": "2021" }, { "id": "thesis:14001", "collection": "thesis", "collection_id": "14001", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:11182020-210226861", "type": "thesis", "title": "Large-Scale Intelligent Systems: From Network Dynamics to Optimization Algorithms", "author": [ { "family_name": "Azizan Ruhi", "given_name": "Navid", "orcid": "0000-0002-4299-2963", "clpid": "Azizan-Ruhi-Navid" } ], "thesis_advisor": [ { "family_name": "Wierman", "given_name": "Adam C.", "clpid": "Wierman-A-C" }, { "family_name": "Hassibi", "given_name": "Babak", "orcid": "0000-0002-1375-5838", "clpid": "Hassibi-B" } ], "thesis_committee": [ { "family_name": "Low", "given_name": "Steven H.", "orcid": "0000-0001-6476-3048", "clpid": "Low-Steven-H" }, { "family_name": "Wierman", "given_name": "Adam C.", "clpid": "Wierman-A-C" }, { "family_name": "Hassibi", "given_name": "Babak", "orcid": "0000-0002-1375-5838", "clpid": "Hassibi-B" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "The expansion of large-scale technological systems such as electrical grids, transportation networks, health care systems, telecommunication networks, the Internet (of things), and other societal networks has created numerous challenges and opportunities at the same time. These systems are often not yet as robust, efficient, sustainable, or smart as we would want them to be. Fueled by the massive amounts of data generated by all these systems, and with the recent advances in making sense out of data, there is a strong desire to make them more intelligent. However, developing large-scale intelligent systems is a multifaceted problem, involving several major challenges. First, large-scale systems typically exhibit complex dynamics due to the large number of entities interacting over a network. Second, because the system is composed of many interacting entities, that make decentralized (and often self-interested) decisions, one has to properly design incentives and markets for such systems. Third, the massive computational needs caused by the scale of the system necessitate performing computations in a distributed fashion, which in turn requires devising new algorithms. Finally, one has to create algorithms that can learn from the copious amounts of data and generalize well. This thesis makes several contributions related to each of these four challenges.
\r\n\r\nAnalyzing and understanding the network dynamics exhibited in societal systems is crucial for developing systems that are robust and efficient. In Part I of this thesis, we study one of the most important families of network dynamics, namely, that of epidemics, or spreading processes. Studying such processes is relevant for understanding and controlling the spread of, e.g., contagious diseases among people, ideas or fake news in online social networks, computer viruses in computer networks, or cascading failures in societal networks. We establish several results on the exact Markov chain model and the nonlinear \"mean-field\" approximations for various kinds of epidemics (i.e., SIS, SIRS, SEIRS, SIV, SEIV, and their variants).
\r\n\r\nDesigning incentives and markets for large-scale systems is critical for their efficient operation and ensuring an alignment between the agents' decentralized decisions and the global goals of the system. To that end, in Part II of this thesis, we study these issues in markets with non-convex costs as well as networked markets, which are of vital importance for, e.g., the smart grid. We propose novel pricing schemes for such markets, which satisfy all the desired market properties. We also reveal issues in the current incentives for distributed energy resources, such as renewables, and design optimization algorithms for efficient management of aggregators of such resources.
\r\n\r\nWith the growing amounts of data generated by large-scale systems, and the fact that the data may already be dispersed across many units, it is becoming increasingly necessary to run computational tasks in a distributed fashion. Part III concerns developing algorithms for distributed computation. We propose a novel consensus-based algorithm for the task of solving large-scale systems of linear equations, which is one of the most fundamental problems in linear algebra, and a key step at the heart of many algorithms in scientific computing, machine learning, and beyond. In addition, in order to deal with the issue of heterogeneous delays in distributed computation, caused by slow machines, we develop a new coded computation technique. In both cases, the proposed methods offer significant speed-ups relative to the existing approaches.
\r\n\r\nOver the past decade, deep learning methods have become the most successful learning algorithms in a wide variety of tasks. However, the reasons behind their success (as well as their failures in some respects) are largely unexplained. It is widely believed that the success of deep learning is not just due to the deep architecture of the models, but also due to the behavior of the optimization algorithms, such as stochastic gradient descent (SGD), used for training them. In Part IV of this thesis, we characterize several properties, such as minimax optimality and implicit regularization, of SGD, and more generally, of the family of stochastic mirror descent (SMD). While SGD performs an implicit regularization, this regularization can be effectively controlled using SMD with a proper choice of mirror, which in turn can improve the generalization error.
", "doi": "10.7907/j494-j572", "publication_date": "2021", "thesis_type": "phd", "thesis_year": "2021" }, { "id": "thesis:14148", "collection": "thesis", "collection_id": "14148", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:05142021-163257155", "primary_object_url": { "basename": "nakka_yashwanth_kumar_2021_thesis.pdf", "content": "final", "filesize": 28258884, "license": "other", "mime_type": "application/pdf", "url": "/14148/2/nakka_yashwanth_kumar_2021_thesis.pdf", "version": "v6.0.0" }, "type": "thesis", "title": "Spacecraft Motion Planning and Control under Probabilistic Uncertainty for Coordinated Inspection and Safe Learning", "author": [ { "family_name": "Nakka", "given_name": "Yashwanth Kumar", "orcid": "0000-0001-7897-3644", "clpid": "Nakka-Yashwanth-Kumar" } ], "thesis_advisor": [ { "family_name": "Chung", "given_name": "Soon-Jo", "orcid": "0000-0002-6657-3907", "clpid": "Chung-Soon-Jo" } ], "thesis_committee": [ { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" }, { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Hadaegh", "given_name": "Fred", "clpid": "Fred-Hadaegh" }, { "family_name": "Chung", "given_name": "Soon-Jo", "orcid": "0000-0002-6657-3907", "clpid": "Chung-Soon-Jo" } ], "local_group": [ { "literal": "GALCIT" }, { "literal": "div_eng" } ], "abstract": "During a spacecraft mission design process, engineers often balance the following three criteria: science return, optimality in performance, and safety. Given a science criterion, engineers design the orbit parameters with predefined performance and safety. Often in this approach, the spacecraft has no understanding of the expected outcome or the knowledge of the mission safety criteria. Autonomous science-driven orbit (or goal) selection and planning for safety under uncertainty enable efficient and adaptable missions. To this end, we propose an architecture for information-based guidance and control for coordinated inspection, motion planning and control algorithms for safe and optimal guidance under uncertainty, and architecture for safe exploration.
\r\n\r\nIn the first part of this thesis, we present an architecture for inspection or mapping of a target spacecraft in a low Earth orbit using multiple observer spacecraft. We use an information gain approach to directly consider the trade-off between gathered data and fuel/energy cost. The estimated information gain is a crucial input to the motion planner, which computes orbits and reconfiguration strategies for each of the observers to maximize the information gain from distributed observations of the target spacecraft. The resulting motion trajectories jointly consider observational coverage of the target spacecraft and fuel/energy cost. We validate our architecture in a mission simulation to visually inspect the target spacecraft and on the three degree-of-freedom robotic spacecraft dynamics simulator testbed.
\r\n\r\nIn the second part of the thesis, we present gPC-SCP, Generalized Polynomial Chaos-based Sequential Convex Programming method, to compute a sub-optimal solution for a continuous-time chance-constrained stochastic nonlinear optimal control (SNOC) problem. The approach enables motion planning and control of robotic systems under uncertainty. The proposed method involves two steps. The first step is to derive a deterministic nonlinear optimal control problem (DNOC) with convex constraints that are surrogate to the SNOC by using gPC expansion and the distributionally-robust convex subset of the chance constraints. The second step is to solve the DNOC problem using sequential convex programming (SCP) for trajectory generation and control. We prove that in the unconstrained case, the optimal value of the DNOC converges to that of SNOC asymptotically and that any feasible solution of the constrained DNOC is a feasible solution of the chance-constrained SNOC. We derive a stable stochastic model predictive controller using the gPC-SCP for tracking a potentially unsafe trajectory in the presence of uncertainty. We empirically demonstrate the efficacy of the gPC-SCP method for the following three test cases: 1) collision checking under uncertainty in actuation, 2) collision checking with stochastic obstacles, and 3) safe trajectory tracking under uncertainty in the dynamics and obstacle location by using a receding horizon control approach. We validate the effectiveness of the gPC-SCP method on the robotic spacecraft testbed.
\r\n\r\n\r\nIn the third part of this thesis, we present a new approach for optimal motion planning for safe exploration that integrates the chance-constrained stochastic optimal control with dynamics learning and feedback control. We derive an iterative convex optimization algorithm that solves an Information-cost Stochastic Nonlinear Optimal Control problem (Info-SNOC). The optimization objective encodes control cost for performance and exploration cost for learning, and the safety is incorporated as distributionally robust chance constraints. The dynamics are predicted from a robust regression model that is learned from data. The Info-SNOC algorithm is used to compute a sub-optimal pool of safe motion plans that aid in exploration for learning unknown residual dynamics under safety constraints. A stable feedback controller is used to execute the motion plan and collect data for model learning. We prove the safety of rollout from our exploration method and reduction in uncertainty over epochs, thereby guaranteeing the consistency of our learning method. We validate the effectiveness of Info-SNOC by designing and implementing a pool of safe trajectories for a planar robot. We demonstrate that our approach has a higher success rate in ensuring safety when compared to a deterministic trajectory optimization approach.
", "doi": "10.7907/6329-sf68", "publication_date": "2021", "thesis_type": "phd", "thesis_year": "2021" }, { "id": "thesis:14105", "collection": "thesis", "collection_id": "14105", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:03152021-082447788", "primary_object_url": { "basename": "dpastorm_thesis_deposit.pdf", "content": "final", "filesize": 18431112, "license": "other", "mime_type": "application/pdf", "url": "/14105/1/dpastorm_thesis_deposit.pdf", "version": "v6.0.0" }, "type": "thesis", "title": "Towards Learning Robotic Dynamics: Application to Multirotor Takeoff and Landing", "author": [ { "family_name": "Pastor", "given_name": "Daniel", "orcid": "0000-0003-3452-0605", "clpid": "Pastor-Daniel" } ], "thesis_advisor": [ { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" } ], "thesis_committee": [ { "family_name": "Chung", "given_name": "Soon-Jo", "orcid": "0000-0002-6657-3907", "clpid": "Chung-Soon-Jo" }, { "family_name": "Gharib", "given_name": "Morteza", "orcid": "0000-0003-0754-4193", "clpid": "Gharib-M" }, { "family_name": "Pellegrino", "given_name": "Sergio", "orcid": "0000-0001-9373-3278", "clpid": "Pellegrino-S" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" } ], "local_group": [ { "literal": "GALCIT" }, { "literal": "div_eng" } ], "abstract": "Multirotors have become widespread but their usage is still limited. Ensuring safety during take-off and landing is still an open problem. Towards this goal this thesis proposes two different solutions to address this problem. The two approaches complement each other and they are tested on hardware.
\r\n \r\nThe first approach is to design a vehicle that is stable during take-off, despite hardware failures or unsteady take-off platforms. A solution is to use a ballistic launch to impose a deterministic path, preventing collisions with its environment. Following this approach led to the development of several SQUID (Streamlined Quick Unfolding Investigation Drone) vehicles. The main challenges are the ballistic initial flight, large accelerations during launch, and limited volume. A first prototype was developed, which is able to transition mid-flight from stable ballistic flight to a fully controllable multirotor. The system has been fabricated and field tested from a moving vehicle up to 50mph to successfully demonstrate the feasibility of the concept and experimentally validate the design's aerodynamic stability and deployment reliability. A second prototype expanded the first one's capabilities incorporating fully-autonomous vision-based navigation, while keeping the ballistic passive stability and stable transition abilities. The new design includes a more reliable plate-based structure and more effective folding fins.
\r\n \r\nThe second approach focuses on designing controllers that are safe regardless of the platform. For that purpose, a Model Predictive Control (MPC) is used to ensure state and input constraints. Given the highly non-linear dynamics platforms and fast dynamics that require a quick controller evaluation, the work in this thesis is built using Koopman Operator theory, which allows tools from linear analysis to be applied to systems with inherently non-linear dynamics. One of the main contributions is a novel method to find Koopman Eigenfunctions directly from data. Another key contribution is an episodic approach to model non-linear actuation dynamics. The proposed method is first tested on simulation and it outperforms comparable approaches. The method is also demonstrated on-board a multirotor for a fast landing application, where the nonlinear ground effect is learned and used to improve landing speed and quality. An additional extension considers model uncertainty in the MPC architecture, where an Ensemble Kalman Sampler is used to learn the uncertainty distribution.
", "doi": "10.7907/199j-dk87", "publication_date": "2021-06-11", "thesis_type": "phd", "thesis_year": "2021" }, { "id": "thesis:14083", "collection": "thesis", "collection_id": "14083", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:02182021-040721884", "primary_object_url": { "basename": "Shi_Xichen_2021_Thesis.pdf", "content": "final", "filesize": 16708714, "license": "other", "mime_type": "application/pdf", "url": "/14083/1/Shi_Xichen_2021_Thesis.pdf", "version": "v5.0.0" }, "type": "thesis", "title": "Intelligent Control for Fixed-Wing eVTOL Aircraft", "author": [ { "family_name": "Shi", "given_name": "Xichen", "orcid": "0000-0002-5366-9256", "clpid": "Shi-Xichen" } ], "thesis_advisor": [ { "family_name": "Chung", "given_name": "Soon-Jo", "orcid": "0000-0002-6657-3907", "clpid": "Chung-Soon-Jo" } ], "thesis_committee": [ { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" }, { "family_name": "Chung", "given_name": "Soon-Jo", "orcid": "0000-0002-6657-3907", "clpid": "Chung-Soon-Jo" }, { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "GALCIT" }, { "literal": "div_eng" } ], "abstract": "Urban Air Mobility (UAM) holds promise for personal air transportation by deploying \"flying cars\" over cities. As such, fixed-wing electric vertical take-off and landing (eVTOL) aircraft has gained popularity as they can swiftly traverse cluttered areas, while also efficiently covering longer distances. These modes of operation call for an enhanced level of precision, safety, and intelligence for flight control. The hybrid nature of these aircraft poses a unique challenge that stems from complex aerodynamic interactions between wings, rotors, and the environment. Thus accurate estimation of external forces is indispensable for a high performance flight. However, traditional methods that stitch together different control schemes often fall short during hybrid flight modes. On the other hand, learning-based approaches circumvent modeling complexities, but they often lack theoretical guarantees for stability.
\r\n\r\nIn the first part of this thesis, we study the theoretical benefits of these fixed-wing eVTOL aircraft, followed by the derivation of a novel unified control framework. It consists of nonlinear position and attitude controllers using forces and moments as inputs; and control allocation modules that determine desired attitudes and thruster signals. Next, we present a composite adaptation scheme for linear-in-parameter (LiP) dynamics models, which provides accurate realtime estimation for wing and rotor forces based on measurements from a three-dimensional airflow sensor. Then, we introduce a design method to optimize multirotor configuration that ensures a property of robustness against rotor failures.
\r\n\r\nIn the second part of the thesis, we use deep neural networks (DNN) to learn part of unmodeled dynamics of the flight vehicles. Spectral normalization that regulates the Lipschitz constants of the neural network is applied for better generalization outside the training domain. The resultant network is utilized in a nonlinear feedback controller with a contraction mapping update, solving the nonaffine-in-control issue that arises. Next, we formulate general methods for designing and training DNN-based dynamics, controller, and observer. The general framework can theoretically handle any nonlinear dynamics with prior knowledge of its structure. Finally, we establish a delay compensation technique that transforms nominal controllers for an undelayed system into a sample-based predictive controller with numerical integration. The proposed method handles both first-order and transport delays in actuators and balances between numerical accuracy and computational efficiency to guarantee stability under strict hardware limitations.
", "doi": "10.7907/51c6-aa57", "publication_date": "2021", "thesis_type": "phd", "thesis_year": "2021" }, { "id": "thesis:13856", "collection": "thesis", "collection_id": "13856", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:08162020-233437139", "primary_object_url": { "basename": "london_palma_2020.pdf", "content": "final", "filesize": 2181183, "license": "other", "mime_type": "application/pdf", "url": "/13856/1/london_palma_2020.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Frameworks for High Dimensional Convex Optimization", "author": [ { "family_name": "London", "given_name": "Palma Alise den Nijs", "orcid": "0000-0001-6472-8293", "clpid": "London-Palma-Alise-den-Nijs" } ], "thesis_advisor": [ { "family_name": "Wierman", "given_name": "Adam C.", "clpid": "Wierman-A-C" } ], "thesis_committee": [ { "family_name": "Low", "given_name": "Steven H.", "orcid": "0000-0001-6476-3048", "clpid": "Low-S-H" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Anandkumar", "given_name": "Anima", "orcid": "0000-0002-6974-6797", "clpid": "Anandkumar-A" }, { "family_name": "Wierman", "given_name": "Adam C.", "clpid": "Wierman-A-C" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "We present novel, efficient algorithms for solving extremely large optimization problems. A significant bottleneck today is that as the size of datasets grow, researchers across disciplines desire to solve prohibitively massive optimization problems. In this thesis, we present methods to compress optimization problems. The general goal is to represent a huge problem as a smaller problem or set of smaller problems, while still retaining enough information to ensure provable guarantees on solution quality and run time. We apply this approach to the following three settings.
\r\n\r\nFirst, we propose a framework for accelerating both linear program solvers and convex solvers for problems with linear constraints. Our focus is on a class of problems for which data is either very costly, or hard to obtain. In these situations, the number of data points m available is much smaller than the number of variables, n. In a machine learning setting, this regime is increasingly prevalent since it is often advantageous to consider larger and larger feature spaces, while not necessarily obtaining proportionally more data. Analytically, we provide worst-case guarantees on both the runtime and the quality of the solution produced. Empirically, we show that our framework speeds up state-of-the-art commercial solvers by two orders of magnitude, while maintaining a near-optimal solution.
\r\n\r\nSecond, we propose a novel approach for distributed optimization which uses far fewer messages than existing methods. We consider a setting in which the problem data are distributed over the nodes. We provide worst-case guarantees on the performance with respect to the amount of communication it requires and the quality of the solution. The algorithm uses O(log(n+m)) messages with high probability. We note that this is an exponential reduction compared to the O(n) communication required during each round of traditional consensus based approaches. In terms of solution quality, our algorithm produces a feasible, near optimal solution. Numeric results demonstrate that the approximation error matches that of ADMM in many cases, while using orders-of-magnitude less communication.
\r\n\r\nLastly, we propose and analyze a provably accurate long-step infeasible Interior Point Algorithm (IPM) for linear programming. The core computational bottleneck in IPMs is the need to solve a linear system of equations at each iteration. We employ sketching techniques to make the linear system computation lighter, by handling well-known ill-conditioning problems that occur when using iterative solvers in IPMs for LPs. In particular, we propose a preconditioned Conjugate Gradient iterative solver for the linear system. Our sketching strategy makes the condition number of the preconditioned system provably small. In practice we demonstrate that our approach significantly reduces the condition number of the linear system, and thus allows for more efficient solving on a range of benchmark datasets.
\r\n", "doi": "10.7907/db29-am33", "publication_date": "2021", "thesis_type": "phd", "thesis_year": "2021" }, { "id": "thesis:14046", "collection": "thesis", "collection_id": "14046", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:01052021-195655093", "primary_object_url": { "basename": "Caltech_Thesis_Richard.pdf", "content": "final", "filesize": 5654612, "license": "other", "mime_type": "application/pdf", "url": "/14046/1/Caltech_Thesis_Richard.pdf", "version": "v5.0.0" }, "type": "thesis", "title": "Assuring Safety under Uncertainty in Learning-Based Control Systems", "author": [ { "family_name": "Cheng", "given_name": "Richard", "orcid": "0000-0001-8301-9169", "clpid": "Cheng-Richard" } ], "thesis_advisor": [ { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" } ], "thesis_committee": [ { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" }, { "family_name": "Ames", "given_name": "Aaron D.", "orcid": "0000-0003-0848-3177", "clpid": "Ames-A-D" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Burdick", "given_name": "Joel Wakeman", "orcid": "0000-0002-3091-540X", "clpid": "Burdick-J-W" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Learning-based controllers have recently shown impressive results for different robotic tasks in well-defined environments, successfully solving a Rubiks cube and sorting objects in a bin. These advancements promise to enable a host of new capabilities for complex robotic systems. However, these learning-based controllers cannot yet be deployed in highly uncertain environments due to significant issues relating to learning reliability, robustness, and safety.
\r\n\r\nTo overcome these issues, this thesis proposes new methods for integrating model information (e.g. model-based control priors) into the reinforcement learning framework, which is crucial to ensuring reliability and safety. I show, both empirically and theoretically, that this model information greatly reduces variance in learning and can effectively constrain the policy search space, thus enabling significant improvements in sample complexity for the underlying RL algorithms. Furthermore, by leveraging control barrier functions and Gaussian process uncertainty models, I show how system safety can be maintained under uncertainty without interfering with the learning process (e.g. distorting the policy gradients).
\r\n\r\nThe last part of the thesis will discuss fundamental limitations that arise when utilizing machine learning to derive safety guarantees. In particular, I show that widely used uncertainty models can be highly inaccurate when predicting rare events, and examine the implications of this for safe learning. To overcome some of these limitations, a novel framework is developed based on assume-guarantee contracts in order to ensure safety in multi-agent human environments. The proposed approach utilizes contracts to impose loose responsibilities on agents in the environment, which are learned from data. Imposing these responsibilities on agents, rather than treating their uncertainty as a purely random process, allows us to achieve both safety and efficiency in interactions.
", "doi": "10.7907/9kye-rn93", "publication_date": "2021", "thesis_type": "phd", "thesis_year": "2021" }, { "id": "thesis:13669", "collection": "thesis", "collection_id": "13669", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:04082020-095943405", "primary_object_url": { "basename": "Cvitkovic_Milan_2020_Thesis.pdf", "content": "final", "filesize": 5963181, "license": "other", "mime_type": "application/pdf", "url": "/13669/1/Cvitkovic_Milan_2020_Thesis.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Deep Learning in Unconventional Domains", "author": [ { "family_name": "Cvitkovic", "given_name": "Michael William (Milan0", "orcid": "0000-0003-4188-452X", "clpid": "Cvitkovic-Michael-William-Milan" } ], "thesis_advisor": [ { "family_name": "Vidick", "given_name": "Thomas G.", "orcid": "0000-0002-6405-365X", "clpid": "Vidick-T" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" }, { "family_name": "Vidick", "given_name": "Thomas G.", "orcid": "0000-0002-6405-365X", "clpid": "Vidick-T" }, { "family_name": "Anandkumar", "given_name": "Anima", "orcid": "0000-0002-6974-6797", "clpid": "Anandkumar-A" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Machine learning methods have dramatically improved in recent years thanks to advances in deep learning (LeCun et al., 2015), a set of methods for training high-dimensional, highly-parameterized, nonlinear functions. Yet deep learning progress has been concentrated in the domains of computer vision, vision-based reinforcement learning, and natural language processing. This dissertation is an attempt to extend deep learning into domains where it has thus far had little impact or has never been applied. It presents new deep learning algorithms and state-of-the-art results on tasks in the domains of source-code analysis, relational databases, and tabular data.
", "doi": "10.7907/v7dm-6r52", "publication_date": "2020", "thesis_type": "phd", "thesis_year": "2020" }, { "id": "thesis:13713", "collection": "thesis", "collection_id": "13713", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:05212020-155425112", "type": "thesis", "title": "Vision for Social Robots: Human Perception and Pose Estimation", "author": [ { "family_name": "Ronchi", "given_name": "Matteo Ruggero", "orcid": "0000-0002-4277-3314", "clpid": "Ronchi-Matteo-Ruggero" } ], "thesis_advisor": [ { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" }, { "family_name": "Ames", "given_name": "Aaron D.", "orcid": "0000-0003-0848-3177", "clpid": "Ames-A-D" }, { "family_name": "Bouman", "given_name": "Katherine L.", "orcid": "0000-0003-0077-4367", "clpid": "Bouman-K-L" }, { "family_name": "Papon", "given_name": "Jeremie", "clpid": "Papon-Jeremie" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "In order to extract the underlying meaning from a scene captured from the surrounding world in a single still image, social robots will need to learn the human ability to detect different objects, understand their arrangement and relationships relative both to their own parts and to each other, and infer the dynamics under which they are evolving. Furthermore, they will need to develop and hold a notion of context to allow assigning different meanings (semantics) to the same visual configuration (syntax) of a scene.
\r\n\r\nThe underlying thread of this Thesis is the investigation of new ways for enabling interactions between social robots and humans, by advancing the visual perception capabilities of robots when they process images and videos in which humans are the main focus of attention.
\r\n\r\nFirst, we analyze the general problem of scene understanding, as social robots moving through the world need to be able to interpret scenes without having been assigned a specific preset goal. Throughout this line of research, i) we observe that human actions and interactions which can be visually discriminated from an image follow a very heavy-tailed distribution; ii) we develop an algorithm that can obtain a spatial understanding of a scene by only using cues arising from the effect of perspective on a picture of a person\u2019s face; and iii) we define a novel taxonomy of errors for the task of estimating the 2D body pose of people in images to better explain the behavior of algorithms and highlight their underlying causes of error.
\r\n\r\nSecond, we focus on the specific task of 3D human pose and motion estimation from monocular 2D images using weakly supervised training data, as accurately predicting human pose will open up the possibility of richer interactions between humans and social robots. We show that when 3D ground-truth data is only available in small quantities, or not at all, it is possible to leverage knowledge about the physical properties of the human body, along with additional constraints related to alternative types of supervisory signals, to learn models that can regress the full 3D pose of the human body and predict its motions from monocular 2D images.
\r\n\r\nTaken in its entirety, the intent of this Thesis is to highlight the importance of, and provide novel methodologies for, social robots' ability to interpret their surrounding environment, learn in a way that is robust to low data availability, and generalize previously observed behaviors to unknown situations in a similar way to humans.
", "doi": "10.7907/n2v1-1g79", "publication_date": "2020", "thesis_type": "phd", "thesis_year": "2020" }, { "id": "thesis:13809", "collection": "thesis", "collection_id": "13809", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06092020-121556493", "primary_object_url": { "basename": "Le_Hoang_2020.pdf", "content": "final", "filesize": 8214263, "license": "other", "mime_type": "application/pdf", "url": "/13809/8/Le_Hoang_2020.pdf", "version": "v5.0.0" }, "type": "thesis", "title": "New Frameworks for Structured Policy Learning", "author": [ { "family_name": "Le", "given_name": "Hoang Minh", "orcid": "0000-0002-5521-5856", "clpid": "Le-Hoang-Minh" } ], "thesis_advisor": [ { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "thesis_committee": [ { "family_name": "Wierman", "given_name": "Adam C.", "orcid": "0000-0002-5923-0199", "clpid": "Wierman-A-C" }, { "family_name": "Anandkumar", "given_name": "Anima", "orcid": "0000-0002-6974-6797", "clpid": "Anandkumar-A" }, { "family_name": "Daum\u00e9", "given_name": "Hal, III", "orcid": "0000-0002-3760-345X", "clpid": "Daum\u00e9-Hal-III" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Sequential decision making applications are playing an increasingly important role in everyday life. Research interest in machine learning approaches to sequential decision making has surged thanks to recent empirical successes of reinforcement learning and imitation learning techniques, partly fueled by recent advances in deep learning-based function approximation. However in many real-world sequential decision making applications, relying purely on black box policy learning is often insufficient, due to practical requirements of data efficiency, interpretability, safety guarantees, etc. These challenges collectively make it difficult for many existing policy learning methods to find success in realistic applications.
\r\n\r\nIn this dissertation, we present recent advances in structured policy learning, which are new machine learning frameworks that integrate policy learning with principled notions of domain knowledge, which spans value-based, policy-based, and model-based structures. Our framework takes flexible reduction-style approaches that can integrate structure with reinforcement learning, imitation learning and robust control techniques. In addition to methodological advances, we demonstrate several successful applications of the new policy learning frameworks.
", "doi": "10.7907/rz4w-k233", "publication_date": "2020", "thesis_type": "phd", "thesis_year": "2020" }, { "id": "thesis:13808", "collection": "thesis", "collection_id": "13808", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06092020-120629159", "type": "thesis", "title": "From Restoring Human Vision to Enhancing Computer Vision", "author": [ { "family_name": "Liu", "given_name": "Yang", "orcid": "0000-0002-8155-9134", "clpid": "Liu-Yang-CNS" } ], "thesis_advisor": [ { "family_name": "Meister", "given_name": "Markus", "clpid": "Meister-M" } ], "thesis_committee": [ { "family_name": "Perona", "given_name": "Pietro", "clpid": "Perona-P" }, { "family_name": "Siapas", "given_name": "Athanassios G.", "clpid": "Siapas-A-G" }, { "family_name": "Yue", "given_name": "Yisong", "clpid": "Yue-Yisong" }, { "family_name": "Meister", "given_name": "Markus", "clpid": "Meister-M" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "The central theme of this work is enabling vision, which includes two subtopics: restoring vision for blind humans, and enhancing computer vision models in visual recognition. Chapter 1 first provides a gentle introduction to relevant high level principles of human visual computations and summarizes two fundamental questions that vision answers: \"what\" and \"where.\" Chapters 2, 3, and 4 contain three published projects that are anchored by those two fundamental questions.
\r\n \r\nChapter 2 introduces a cognitive assistant to restore visual function for blind humans by focusing on an interface powered by audio augmented reality. The assistant communicates the \"what\" and \"where\" aspects of visual scenes by a combination of natural language and spatialized sound. We experimentally demonstrated that the assistant enables many aspects of visual functions for naive blind users.
\r\n \r\nChapters 3 and 4 develop data augmentation methods to address the data inefficiency problem in neural network based computer visual recognition models. In Chapter 3, a 3D-simulation based data augmentation method is developed for improving the generalization of visual classification models for rare classes. In Chapter 4, a fast and efficient data augmentation method is developed for the newly formulated panoptic segmentation task. The method improves performance of state-of-the-art panoptic segmentation models and generalizes across dataset domains, sizes, model architectures, and backbones.
", "doi": "10.7907/sq58-z682", "publication_date": "2020", "thesis_type": "phd", "thesis_year": "2020" }, { "id": "thesis:13755", "collection": "thesis", "collection_id": "13755", "cite_using_url": "https://resolver.caltech.edu/CaltechThesis:06012020-093627645", "primary_object_url": { "basename": "Tokpanov_Yury_2020.pdf", "content": "final", "filesize": 18652320, "license": "other", "mime_type": "application/pdf", "url": "/13755/1/Tokpanov_Yury_2020.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Towards Next Generation of Optoelectronics: from Quantum Plasmonics and 2D Materials to Advanced Optimization Techniques of Nanophotonic Devices", "author": [ { "family_name": "Tokpanov", "given_name": "Yury", "orcid": "0000-0001-5123-7428", "clpid": "Tokpanov-Yury" } ], "thesis_advisor": [ { "family_name": "Atwater", "given_name": "Harry Albert", "clpid": "Atwater-H-A" } ], "thesis_committee": [ { "family_name": "Faraon", "given_name": "Andrei", "clpid": "Faraon-A" }, { "family_name": "Vahala", "given_name": "Kerry J.", "clpid": "Vahala-K-J" }, { "family_name": "Atwater", "given_name": "Harry Albert", "clpid": "Atwater-H-A" }, { "family_name": "Yue", "given_name": "Yisong", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "In this thesis, we explore different novel concepts and materials for the next-generation of nanophotonic and optoelectronic devices that could be used both in classical and quantum settings.
\r\n\r\nFirst, we study quantum coherence properties of surface plasmon polaritons (SPPs) in the regime of extreme dispersion. Most experiments to date, that tested quantum coherence properties of SPPs, used essentially weakly-confined plasmons, which experience limited light-matter hybridization, thus restricting the potential for decoherence. Our setup is based on a hole-array chip supporting SPPs near the surface plasma frequency, where plasmonic dispersion and confinement is much stronger than in previous experiments, making the plasmons much more susceptible for decoherence processes. We generated polarization-entangled pairs of photons and transmitted one of the photons through this plasmonic hole array. Our results show that the quality of photon entanglement after the highly-dispersive plasmonic channel is unperturbed. Our findings provide a lower bound of 100 femtoseconds for the pure dephasing time of dispersive plasmons in our materials, and show that even in a highly dispersive regime, surface plasmons preserve quantum mechanical correlations, making possible harnessing the power of extreme light confinement for integrated quantum photonics.
\r\n\r\nSecond, we systematically study different passivation schemes of sulfur vacancies in 2D molybdenum disulfide using first-principles calculations based on density functional theory. We aim at building a microscopic understanding of passivation mechanisms of treatment with TFSI superacid - a popular approach of to improve optical properties. Since superacids have a strong ability to donate protons, we consider hydrogenation and protonation of sulfur vacancies as a possible passivation scheme. Our calculations show that effects of protonation and hydrogenation on properties of 2D molybdenum disulfide are very similar. Moreover, we find that four hydrogen atoms can fully \"heal\" sulfur vacancies in this material. Our results are an important step towards controllable defects design in 2D transition metal dichalcogenides.
\r\n\r\nAnd third, we study applications of advanced methods of optimization and machine learning to the design of different nanophotonic devices. We explore feasibility of using novel multi-fidelity Gaussian processes optimization technique to optimize plasmonic mirror filters for hyperspectral imaging. We compare our results with other common optimization approaches. Then we apply deep-learning inspired techniques to optimize control voltages of individual pixels of active metasurfaces to achieve dynamic beamsteering. We obtain interesting results that pave the way for future experiments both in nanophotonics and machine learning fields.
", "doi": "10.7907/tg1b-hn35", "publication_date": "2020", "thesis_type": "phd", "thesis_year": "2020" }, { "id": "thesis:11326", "collection": "thesis", "collection_id": "11326", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:12222018-173706714", "type": "thesis", "title": "Probabilistic Protein Engineering", "author": [ { "family_name": "Yang", "given_name": "Kevin Kaichuang", "orcid": "0000-0001-9045-6826", "clpid": "Yang-Kevin-Kaichuang" } ], "thesis_advisor": [ { "family_name": "Arnold", "given_name": "Frances Hamilton", "orcid": "0000-0002-4027-364X", "clpid": "Arnold-F-H" } ], "thesis_committee": [ { "family_name": "Clemons", "given_name": "William M.", "orcid": "0000-0002-0021-889X", "clpid": "Clemons-W-M" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Tirrell", "given_name": "David A.", "orcid": "0000-0003-3175-4596", "clpid": "Tirrell-D-A" }, { "family_name": "Arnold", "given_name": "Frances Hamilton", "orcid": "0000-0002-4027-364X", "clpid": "Arnold-F-H" } ], "local_group": [ { "literal": "Rosen Bioengineering Center" }, { "literal": "div_chem" } ], "abstract": "Machine learning-guided protein engineering is a new paradigm that enables the optimization of complex protein functions. Machine-learning methods use data to predict protein function without requiring a detailed model of the underlying physics or biological pathways. They accelerate protein engineering by learning from information contained in all measured variants and using it to select variants that are likely to be improved. We begin with a review of the basics of machine learning with a focus on applications to protein engineering and protein sequence-function datasets (Chapter 1). We used the entire machine-learning guided engineering paradigm to engineer the algal-derived light-gated channel channelrhodopsin (ChR), which can be used to modulate neuronal activity with light. We build models that discover ChRs with strong plasma membrane localization in mammalian cells (Chapter 2) and unprecedented light sensitivity and photocurrents for optogenetic applications (Chapter 3). Machine learning-guided evolution requires a machine-learning model that learns the relationship between sequence and function. For machine-learning models to learn about protein sequences, protein sequences must be represented as vectors or matrices of numbers. How each protein sequence is represented determines what can be learned. We learn continuous vector encodings of sequences from patterns in unlabeled sequences (Chapter 4). Learned encodings are low-dimensional, do not require alignments, and may improve performance by transferring information in unlabeled sequences to specific prediction tasks. Alternately, we demonstrate an interpretable Gaussian process kernel tailored to biological sequences (Chapter 6). In addition to a model to predict function from sequence, engineering requires a method to use the model to choose sequences for the next round of evolution. Most machine-learning guided engineering strategies assume that selected sequences can be queried directly. However, in directed evolution it is common to design a library of sequences and then sample stochastic batches from that library. We propose a batched stochastic Bayesian optimization algorithm for iteratively designing and screening site-saturation mutagenesis libraries (Chapter 5).
", "doi": "10.7907/ZR4M-K630", "publication_date": "2019", "thesis_type": "phd", "thesis_year": "2019" }, { "id": "thesis:11502", "collection": "thesis", "collection_id": "11502", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:05082019-103122440", "primary_object_url": { "basename": "Towards_a_Visipedia__Tools_and_Techniques_for_Computer_Vision_Dataset_Collection (4).pdf", "content": "final", "filesize": 20651754, "license": "other", "mime_type": "application/pdf", "url": "/11502/1/Towards_a_Visipedia__Tools_and_Techniques_for_Computer_Vision_Dataset_Collection (4).pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Towards a Visipedia: Combining Computer Vision and Communities of Experts", "author": [ { "family_name": "Van Horn", "given_name": "Grant Richard", "orcid": "0000-0003-2953-9651", "clpid": "Van-Horn-Grant-Richard" } ], "thesis_advisor": [ { "family_name": "Perona", "given_name": "Pietro", "clpid": "Perona-P" } ], "thesis_committee": [ { "family_name": "Yue", "given_name": "Yisong", "clpid": "Yue-Yisong" }, { "family_name": "Wierman", "given_name": "Adam C.", "clpid": "Wierman-A-C" }, { "family_name": "Belongie", "given_name": "Serge J.", "clpid": "Belongie-S-J" }, { "family_name": "Perona", "given_name": "Pietro", "clpid": "Perona-P" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Motivated by the idea of a Visipedia, where users can search and explore by image, this thesis presents tools and techniques for empowering expert communities through computer vision. The collective aim of this work is to provide a scalable foundation upon which an application like Visipedia can be built. We conduct experiments using two highly motivated communities, the birding community and the naturalist community, and report results and lessons on how to build the necessary components of a Visipedia. First, we conduct experiments analyzing the behavior of state-of-the-art computer vision classifiers on long tailed datasets. We find poor feature sharing between classes, potentially limiting the applicability of these models and emphasizing the ability to intelligently direct data collection resources. Second, we devise online crowdsourcing algorithms to make dataset collection for binary labels, multiclass labels, keypoints, and mulit-instance bounding boxes faster, cheaper, and more accurate. These methods jointly estimate labels, worker skills, and train computer vision models for these tasks. Experiments show that we can achieve significant cost savings compared to traditional data collection techniques, and that we can produce a more accurate dataset compared to traditional data collection techniques. Third, we present two fine-grained datasets, detail how they were constructed, and analyze the test accuracy of state-of-the-art methods. These datasets are then used to create applications that help users identify species in their photographs: Merlin, an app assisting users in identifying birds species, and iNaturalist, an app that assists users in identifying a broad variety of species. Finally, we present work aimed at reducing the computational burden of large scale classification with the goal of creating an application that allows users to classify tens of thousands of species in real time on their mobile device. As a whole, the lessons learned and the techniques presented in this thesis bring us closer to the realization of a Visipedia.
", "doi": "10.7907/20DQ-Y220", "publication_date": "2019", "thesis_type": "phd", "thesis_year": "2019" }, { "id": "thesis:10978", "collection": "thesis", "collection_id": "10978", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:05312018-080301508", "primary_object_url": { "basename": "Ren_Xiaoqi_2018.pdf", "content": "final", "filesize": 6693722, "license": "other", "mime_type": "application/pdf", "url": "/10978/1/Ren_Xiaoqi_2018.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Optimizing Resource Management in Cloud Analytics Services", "author": [ { "family_name": "Ren", "given_name": "Xiaoqi", "orcid": "0000-0002-1121-9046", "clpid": "Ren-Xiaoqi" } ], "thesis_advisor": [ { "family_name": "Wierman", "given_name": "Adam C.", "clpid": "Wierman-A-C" } ], "thesis_committee": [ { "family_name": "Wierman", "given_name": "Adam C.", "clpid": "Wierman-A-C" }, { "family_name": "Low", "given_name": "Steven H.", "clpid": "Low-S-H" }, { "family_name": "Chandy", "given_name": "K. Mani", "clpid": "Chandy-K-M" }, { "family_name": "Yue", "given_name": "Yisong", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "Resnick Sustainability Institute" }, { "literal": "div_eng" } ], "abstract": "The fundamental challenge in the cloud today is how to build and optimize machine learning and data analytical services. Machine learning and data analytical platforms are changing computing infrastructure from expensive private data centers to easily accessible online services. These services pack user requests as jobs and run them on thousands of machines in parallel in geo-distributed clusters. The scale and the complexity of emerging jobs lead to increasing challenges for the clusters at all levels, from power infrastructure to system architecture and corresponding software framework design.
\r\n\r\nThese challenges come in many forms. Today's clusters are built on commodity hardware and hardware failures are unavoidable. Resource competition, network congestion, and mixed generations of hardware make the hardware environment complex and hard to model and predict. Such heterogeneity becomes a crucial roadblock for efficient parallelization on both the task level and job level. Another challenge comes from the increasing complexity of the applications. For example, machine learning services run jobs made up of multiple tasks with complex dependency structures. This complexity leads to difficulties in framework designs. The scale, especially when services span geo-distributed clusters, leads to another important hurdle for cluster design. Challenges also come from the power infrastructure. Power infrastructure is very expensive and accounts for more than 20% of the total costs to build a cluster. Power sharing optimization to maximize the facility utilization and smooth peak hour usages is another roadblock for cluster design.
\r\n\r\nIn this thesis, we focus on solutions for these challenges at the task level, on the job level, with respect to the geo-distributed data cloud design and for power management in colocation data centers.
\r\n\r\nAt the task level, a crucial hurdle to achieving predictable performance is stragglers, i.e., tasks that take significantly longer than expected to run. At this point, speculative execution has been widely adopted to mitigate the impact of stragglers in simple workloads. We apply straggler mitigation for approximation jobs for the first time. We present GRASS, which carefully uses speculation to mitigate the impact of stragglers in approximation jobs. GRASS's design is based on the analysis of a model we develop to capture the optimal speculation levels for approximation jobs. Evaluations with production workloads from Facebook and Microsoft Bing in an EC2 cluster of 200 nodes show that GRASS increases accuracy of deadline-bound jobs by 47% and speeds up error-bound jobs by 38%.
\r\n\r\nMoving from task level to job level, task level speculation mechanisms are designed and operated independently of job scheduling when, in fact, scheduling a speculative copy of a task has a direct impact on the resources available for other jobs. Thus, we present Hopper, a job-level speculation-aware scheduler that integrates the tradeoffs associated with speculation into job scheduling decisions based on a model generalized from the task-level speculation model. We implement both centralized and decentralized prototypes of the Hopper scheduler and show that 50% (66%) improvements over state-of-the-art centralized (decentralized) schedulers and speculation strategies can be achieved through the coordination of scheduling and speculation.
\r\n\r\nAs computing resources move from local clusters to geo-distributed cloud services, we are expecting the same transformation for data storage. We study two crucial pieces of a geo-distributed data cloud system: data acquisition and data placement. Starting from developing the optimal algorithm for the case of a data cloud made up of a single data center, we propose a near-optimal, polynomial-time algorithm for a geo-distributed data cloud in general. We show, via a case study, that the resulting design, Datum, is near-optimal (within 1.6%) in practical settings.
\r\n \r\nEfficient power management is a fundamental challenge for data centers when providing reliable services. Power oversubscription in data centers is very common and may occasionally trigger an emergency when the aggregate power demand exceeds the capacity. We study power capping solutions for handling such emergencies in a colocation data center, where the operator supplies power to multiple tenants. We propose a novel market mechanism based on supply function bidding, called COOP, to financially incentivize and coordinate tenants' power reduction for minimizing total performance loss while satisfying multiple power capping constraints. We demonstrate that COOP is \"win-win\", increasing the operator's profit (through oversubscription) and reducing tenants' costs (through financial compensation for their power reduction during emergencies).
", "doi": "10.7907/K62Y-FV39", "publication_date": "2018", "thesis_type": "phd", "thesis_year": "2018" }, { "id": "thesis:10447", "collection": "thesis", "collection_id": "10447", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:09222017-130217881", "type": "thesis", "title": "Graph Clustering: Algorithms, Analysis and Query Design", "author": [ { "family_name": "Korlakai Vinayak", "given_name": "Ramya", "orcid": "0000-0003-0248-9551", "clpid": "Korlakai-Vinayak-Ramya" } ], "thesis_advisor": [ { "family_name": "Hassibi", "given_name": "Babak", "clpid": "Hassibi-B" } ], "thesis_committee": [ { "family_name": "Hassibi", "given_name": "Babak", "clpid": "Hassibi-B" }, { "family_name": "Wierman", "given_name": "Adam C.", "clpid": "Wierman-A-C" }, { "family_name": "Perona", "given_name": "Pietro", "clpid": "Perona-P" }, { "family_name": "Chandrasekaran", "given_name": "Venkat", "clpid": "Chandrasekaran-V" }, { "family_name": "Yue", "given_name": "Yisong", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "A wide range of applications in engineering as well as the natural and social sciences have datasets that are unlabeled. Clustering plays a major role in exploring structure in such unlabeled datasets. Owing to the heterogeneity in the applications and the types of datasets available, there are plenty of clustering objectives and algorithms. In this thesis we focus on two such clustering problems: Graph Clustering and Crowdsourced Clustering.
\r\n\r\nIn the first part, we consider the problem of graph clustering and study convex-optimization-based clustering algorithms. Datasets are often messy -- ridden with noise, outliers (items that do not belong to any clusters), and missing data. Therefore, we are interested in algorithms that are robust to such discrepancies. We present and analyze convex-optimization-based clustering algorithms which aim to recover the low-rank matrix that encodes the underlying cluster structure for two clustering objectives: clustering partially observed graphs and clustering similarity matrices with outliers. Using block models as generative models, we characterize the performance of these convex clustering algorithms. In particular, we provide explicit bounds, without any large unknown constants, on the problem parameters that determine the success and failure of these convex approaches.
\r\n\r\nIn the second part, we consider the problem of crowdsourced clustering -- the task of clustering items using answers from non-expert crowd workers who can answer similarity comparison queries. Since the workers are not experts, they provide noisy answers. Further, due to budget constraints, we cannot make all possible comparisons between items in the dataset. Thus, it is important to design queries that can reduce the noise in the responses and design algorithms that can work with noisy and partial data. We demonstrate that random triangle queries (where three items are compared per query) provide less noisy data as well as greater quantity of data, for a fixed query budget, as compared to random edge queries (where two items are compared per query). We extend the analysis of convex clustering algorithms to show that the exact recovery guarantees hold for triangle queries despite involving dependent edges. In addition to random querying strategies, we also present a novel active querying algorithm that is guaranteed to find all the clusters regardless of their sizes and without the knowledge of any parameters as long as the workers are better than random guessers. We also provide a tight upper bound on the number of queries made by the proposed active querying algorithm. Apart from providing theoretical guarantees for the clustering algorithms we also apply our algorithms to real datasets.
", "doi": "10.7907/Z9RR1WFK", "publication_date": "2018", "thesis_type": "phd", "thesis_year": "2018" }, { "id": "thesis:10797", "collection": "thesis", "collection_id": "10797", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:04092018-211419944", "primary_object_url": { "basename": "thesis_final.pdf", "content": "final", "filesize": 8151137, "license": "other", "mime_type": "application/pdf", "url": "/10797/1/thesis_final.pdf", "version": "v5.0.0" }, "type": "thesis", "title": "Measuring R(D(*)) for B \u2192 \u203eD(*)\u03c4\u03bd_\u03c4 using Semileptonic Tags and Tau Decays to Hadrons", "author": [ { "family_name": "Chao", "given_name": "Daniel Shuteh", "clpid": "Chao-Daniel-Shuteh" } ], "thesis_advisor": [ { "family_name": "Porter", "given_name": "Frank C.", "clpid": "Porter-F-C" } ], "thesis_committee": [ { "family_name": "Porter", "given_name": "Frank C.", "clpid": "Porter-F-C" }, { "family_name": "Hitlin", "given_name": "David G.", "clpid": "Hitlin-D-G" }, { "family_name": "Carroll", "given_name": "Sean M.", "clpid": "Carroll-S-M" }, { "family_name": "Yue", "given_name": "Yisong", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_pma" } ], "abstract": "We perform a measurement of R(D(*)) for B \u2192 D(*)\u03c4\u03bd\u03c4 using semileptonic tagging and \u03c4 decays to hadrons on the 429 fb-1 of data that BABAR collected at the \u03a5(4S) resonance. This is the first measurement of R(D(*)) using the specied reconstruction channels. Candidate selection was performed with supervised learning, where the training labels were obtained by solving an instance of subgraph isomorphism. The signal extraction was performed by solving an optimization problem whose objective function required the evaluation of kernel density estimates that were accelerated by a branch-and-bound algorithm as well as with a GPU. The training data for the density estimates were themselves the output of two classier scores. We present a 68% and 95% confidence regions of R(D(*)), which do not show enough evidence to reject the standard model prediction.
", "doi": "10.7907/25e6-j889", "publication_date": "2018", "thesis_type": "phd", "thesis_year": "2018" }, { "id": "thesis:10451", "collection": "thesis", "collection_id": "10451", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:09242017-121200052", "type": "thesis", "title": "Reducing Latencies in Earthquake Early Warning", "author": [ { "family_name": "Yin", "given_name": "Lucy", "orcid": "0000-0002-0652-9330", "clpid": "Yin-Lucy" } ], "thesis_advisor": [ { "family_name": "Heaton", "given_name": "Thomas H.", "orcid": "0000-0003-3363-2197", "clpid": "Heaton-T-H" } ], "thesis_committee": [ { "family_name": "Asimaki", "given_name": "Domniki", "orcid": "0000-0002-3008-8088", "clpid": "Asimaki-D" }, { "family_name": "Ampuero", "given_name": "Jean-Paul", "orcid": "0000-0002-4827-7987", "clpid": "Ampuero-J-P" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" }, { "family_name": "Heaton", "given_name": "Thomas H.", "orcid": "0000-0003-3363-2197", "clpid": "Heaton-T-H" }, { "family_name": "Page", "given_name": "Morgan", "clpid": "Page-Morgan" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Existing Earthquake Early Warning (EEW) algorithms use waveform analysis for earthquake detections, estimation of source parameters (i.e., magnitude and hypocenter location), and prediction of peak ground motions at sites near the source. The latency of warning delivery due to data collection significantly restricts the usefulness of the system, especially for users in the vicinity of the earthquake source, as the warning may not arrive before the strong shaking. This presentation discusses several methods to reduce the warning latency, while maintaining reliability and robustness, so that the warning time can be maximized for users to take appropriate actions to reduce causalities and economic losses.
\r\n\r\nFirstly, we incorporated the seismicity forecast information from Epidemic-Type Aftershock Sequence (ETAS) model into EEW as prior information, under the Bayesian probabilistic inference framework. Similar to human\u2019s decision-making process, the Bayesian approach updates the probability of the estimations as more information becomes available. This allows us to reduce the required time for reliable earthquake signal detection from at least 3 seconds to 0.5 second. Furthermore, the initial error of hypocenter location estimation is reduced by 58%. The performance of the algorithm is further improved during aftershock sequences and swarm earthquakes.
\r\n\r\nSecondly, we introduce the use of multidimensional (KD tree) data structure to organize seismic database, so that the querying time can be reduced for the nearest neighbor search during earthquake source parameter estimation. The processing time of KD tree is approximately 15% of the processing time of linear exhaustive search, which allows the potential use of large seismic databases in real-time.
\r\n\r\nEEW is an interdisciplinary subject that involves collaboration among different scientific and engineering communities. Only by optimizing the warning time, such a unified system could be successful in taking protective actions before, during, and after earthquake natural disasters.
", "doi": "10.7907/Z9TH8JW4", "publication_date": "2018", "thesis_type": "phd", "thesis_year": "2018" }, { "id": "thesis:10530", "collection": "thesis", "collection_id": "10530", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:10182017-210853845", "primary_object_url": { "basename": "Niangjun_Chen_2017.pdf", "content": "final", "filesize": 1603423, "license": "other", "mime_type": "application/pdf", "url": "/10530/1/Niangjun_Chen_2017.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Online Algorithms: From Prediction to Decision\r ", "author": [ { "family_name": "Chen", "given_name": "Niangjun", "orcid": "0000-0002-2289-9737", "clpid": "Chen-Niangjun" } ], "thesis_advisor": [ { "family_name": "Wierman", "given_name": "Adam C.", "clpid": "Wierman-A-C" }, { "family_name": "Low", "given_name": "Steven H.", "clpid": "Low-S-H" } ], "thesis_committee": [ { "family_name": "Wierman", "given_name": "Adam C.", "clpid": "Wierman-A-C" }, { "family_name": "Low", "given_name": "Steven H.", "clpid": "Low-S-H" }, { "family_name": "Chandrasekaran", "given_name": "Venkat", "clpid": "Chandrasekaran-V" }, { "family_name": "Yue", "given_name": "Yisong", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Making use of predictions is a crucial, but under-explored, area of sequential decision problems with limited information. While in practice most online algorithms rely on predictions to make real time decisions, in theory their performance is only analyzed in simplified models of prediction noise, either adversarial or i.i.d. The goal of this thesis is to bridge this divide between theory and practice: to study online algorithm under more practical predictions models, gain better understanding about the value of prediction, and design online algorithms that make the best use of predictions.
\r\n\r\nThis thesis makes three main contributions. First, we propose a stochastic prediction error model that generalizes prior models in the learning and stochastic control communities, incorporates correlation among prediction errors, and captures the fact that predictions improve as time passes. Using this general prediction model, we prove that Averaging Fixed Horizon Control (AFHC) can simultaneously achieve sublinear regret and constant competitive ratio in expectation using only a constant- sized prediction window, overcoming the hardnesss results in adversarial prediction models. Second, to understand the optimal use of noisy prediction, we introduce a new class of policies, Committed Horizon Control (CHC), that generalizes both popular policies Receding Horizon Control (RHC) and Averaging Fixed Horizon Control (AFHC). Our results provide explicit results characterizing the optimal use of prediction in CHC policy as a function of properties of the prediction noise, e.g., variance and correlation structure. Third, we apply the general prediction model and algorithm design framework to the deferrable load control problem in power systems. Our proposed model predictive algorithm provides significant reduction in variance of total load in the power system. Throughout this thesis, we provide both average-case analysis and concentration results for our proposed online algorithms, highlighting that the typical performance is tightly concentrated around the average-case performance.
", "doi": "10.7907/Z95M63W4", "publication_date": "2018", "thesis_type": "phd", "thesis_year": "2018" }, { "id": "thesis:10936", "collection": "thesis", "collection_id": "10936", "cite_using_url": "https://resolver.caltech.edu/CaltechThesis:05252018-092016207", "primary_object_url": { "basename": "Zheng_Stephan_2018.pdf", "content": "final", "filesize": 15518224, "license": "other", "mime_type": "application/pdf", "url": "/10936/1/Zheng_Stephan_2018.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Exploiting Structure for Scalable and Robust Deep Learning", "author": [ { "family_name": "Zheng", "given_name": "Stephan Tao", "orcid": "0000-0002-7271-1616", "clpid": "Zheng-Stephan-Tao" } ], "thesis_advisor": [ { "family_name": "Yue", "given_name": "Yisong", "clpid": "Yue-Yisong" } ], "thesis_committee": [ { "family_name": "Perona", "given_name": "Pietro", "clpid": "Perona-P" }, { "family_name": "Porter", "given_name": "Frank C.", "clpid": "Porter-F-C" }, { "family_name": "Sutskever", "given_name": "Ilya", "clpid": "Sutskever-Ilya" }, { "family_name": "Yue", "given_name": "Yisong", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_pma" } ], "abstract": "Deep learning has seen great success training deep neural networks for complex prediction problems, such as large-scale image recognition, short-term time-series forecasting, and learning behavioral models for games with simple dynamics. However, neural networks have a number of weaknesses: 1) they are not sample-efficient and 2) they are often not robust against (adversarial) input perturbations. Hence, it is challenging to train neural networks for problems with exponential complexity, such as multi-agent games, complex long-term spatiotemporal dynamics, or noisy high-resolution image data.
\r\n\r\nThis thesis contributes methods to improve the sample efficiency, expressive power, and robustness of neural networks, by exploiting various forms of low-dimensional structure, such as spatiotemporal hierarchy and multi-agent coordination. We show the effectiveness of this approach in multiple learning paradigms: in both the supervised learning (e.g., imitation learning) and reinforcement learning settings.
\r\n\r\nFirst, we introduce hierarchical neural networks that model both short-term actions and long-term goals from data, and can learn human-level behavioral models for spatiotemporal multi-agent games, such as basketball, using imitation learning.
\r\n\r\nSecond, in reinforcement learning, we show that behavioral policies with a hierarchical latent structure can efficiently learn forms of multi-agent coordination, which enables a form of structured exploration for faster learning.
\r\n\r\nThird, we showcase tensor-train recurrent neural networks that can model high-order mutliplicative structure in dynamical systems (e.g., Lorenz dynamics). We show that this model class gives state-of-the-art long-term forecasting performance with very long time horizons for both simulation and real-world traffic and climate data.
\r\n\r\nFinally, we demonstrate two methods for neural network robustness: 1) stability training, a form of stochastic data augmentation to make neural networks more robust, and 2) neural fingerprinting, a method that detects adversarial examples by validating the network\u2019s behavior in the neighborhood of any given input.
\r\n\r\nIn sum, this thesis takes a step to enable machine learning for the next scale of problem complexity, such as rich spatiotemporal multi-agent games and large-scale robust predictions.
", "doi": "10.7907/4S2Y-CY80", "publication_date": "2018", "thesis_type": "phd", "thesis_year": "2018" }, { "id": "thesis:10281", "collection": "thesis", "collection_id": "10281", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06052017-121818057", "type": "thesis", "title": "Computational Methods for Behavior Analysis", "author": [ { "family_name": "Eyjolfsdottir", "given_name": "Eyrun-Arna", "clpid": "Eyjolfsdottir-Eyrun-Arna" } ], "thesis_advisor": [ { "family_name": "Perona", "given_name": "Pietro", "clpid": "Perona-P" } ], "thesis_committee": [ { "family_name": "Perona", "given_name": "Pietro", "clpid": "Perona-P" }, { "family_name": "Yue", "given_name": "Yisong", "clpid": "Yue-Yisong" }, { "family_name": "Anderson", "given_name": "David J.", "clpid": "Anderson-D-J" }, { "family_name": "Fowlkes", "given_name": "Charless C.", "clpid": "Fowlkes-C-C" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Behavioral scientists strive to decode the functional relationship between sensory input and motor output of the brain, which requires quantitive measurement of animal behavior. Artificial intelligence researchers aim to build intelligent systems, capable of understanding, predicting, and generating behavior. Our research lies on the intersection of the two fields; our goal is to automate measurement of animal behavior and to model their sensory-motor relationship using machine learning.
\r\n\r\nWe have developed a tool that tracks the pose of multiple fruit flies and aims to maintain their identity throughout a video. It outputs motion trajectories that can be used to quantify behavioral differences between individuals, for example by comparing histograms of velocities and wing angles. We show that the tool also works well on non-fly-like animals such as zebrafish larvae.
\r\n\r\nEmbedded in these motion trajectories are temporal patterns that constitute actions. We developed two supervised learning frameworks for action detection: a sliding window framework and a structured output framework. Both frameworks learn to classify actions from motion trajectories and expert annotated action intervals. Our results show that the simpler sliding window framework achievers better results in spite of being much faster to train, reaching 90% of human performance.
\r\n\r\nSupervised learning requires a lot of training data which involves time consuming and painstaking annotation. To alleviate that we have built a semi-supervised neural network framework that, in addition to classifying actions, learns to predict how an animal will move next given its motion and sensory inputs so far. Our model archives as good results as its supervised counterpart with only half of the expert labels. \r\nIn addition, we show that motion prediction can be used to generate convincing simulations of fruit fly behavior and handwritten text, and that our model learns to represent high level information, such as identity, when trained unsupervised.
\r\n\r\nAlthough developed for animal behavior, our methods are general and could be applied to other motion data. We hope that this thesis demonstrates the value of studying animal behavior for the development of artificial intelligence.
", "doi": "10.7907/Z9445JJH", "publication_date": "2017", "thesis_type": "phd", "thesis_year": "2017" }, { "id": "thesis:9987", "collection": "thesis", "collection_id": "9987", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:11292016-204528802", "type": "thesis", "title": "Automated Macro-scale Causal Hypothesis Formation Based on Micro-scale Observation", "author": [ { "family_name": "Chalupka", "given_name": "Krzysztof", "clpid": "Chalupka-Krzysztof" } ], "thesis_advisor": [ { "family_name": "Perona", "given_name": "Pietro", "clpid": "Perona-P" }, { "family_name": "Eberhardt", "given_name": "Frederick", "clpid": "Eberhardt-F" } ], "thesis_committee": [ { "family_name": "Perona", "given_name": "Pietro", "clpid": "Perona-P" }, { "family_name": "Eberhardt", "given_name": "Frederick D.", "clpid": "Eberhardt-F-D" }, { "family_name": "Yue", "given_name": "Yisong", "clpid": "Yue-Yisong" }, { "family_name": "Tsao", "given_name": "Doris Y.", "clpid": "Tsao-D-Y" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "This book introduces new concepts at the intersection of machine learning, causal inference and philosophy of science: the macrovariable cause and effect. Methods for learning such from microvariable data are introduced. The learning process proposes a minimal number of guided experiments that recover the macrovariable cause from observational data.
\r\n\r\nMathematical definitions of a micro- and macro- scale manipulation, an observational and causal partition, and a subsidiary variable are given. These concepts provide a link to previous work in causal inference and machine learning.
\r\n\r\nThe main theoretical result is the Causal Coarsening Theorem, a new insight into the measure-theoretic structure of probability spaces and structural equation models. The theorem provides grounds for automatic causal hypothesis formation from data. Other results concern the minimality and sufficiency of representations created in accordance with the theorem.
\r\n\r\nFinally, this book proposes the first algorithms for supervised and unsupervised causal macrovariable discovery. These algorithms bridge large-scale, multidimensional machine learning and causal inference. In an application to climate science, the algorithms re-discover a known causal mechanism as a viable causal hypothesis. In a psychophysical experiment, the algorithms learn to minimally change visual stimuli to achieve a desired effect on human perception.
", "doi": "10.7907/Z9MW2F4P", "publication_date": "2017", "thesis_type": "phd", "thesis_year": "2017" }, { "id": "thesis:10138", "collection": "thesis", "collection_id": "10138", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:04172017-163725367", "primary_object_url": { "basename": "online-learning-control.pdf", "content": "final", "filesize": 12028711, "license": "other", "mime_type": "application/pdf", "url": "/10138/1/online-learning-control.pdf", "version": "v2.0.0" }, "type": "thesis", "title": "Online Learning for the Control of Human Standing via Spinal Cord Stimulation", "author": [ { "family_name": "Sui", "given_name": "Yanan", "orcid": "0000-0002-9480-627X", "clpid": "Sui-Yanan" } ], "thesis_advisor": [ { "family_name": "Burdick", "given_name": "Joel Wakeman", "clpid": "Burdick-J-W" } ], "thesis_committee": [ { "family_name": "Burdick", "given_name": "Joel Wakeman", "clpid": "Burdick-J-W" }, { "family_name": "Murray", "given_name": "Richard M.", "clpid": "Murray-R-M" }, { "family_name": "Perona", "given_name": "Pietro", "clpid": "Perona-P" }, { "family_name": "Yue", "given_name": "Yisong", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Many applications in recommender systems or experimental design need to make decisions online. Each decision leads to a stochastic reward with initially unknown distribution, while new decisions are made based on the observations of previous rewards. To maximize the total reward, one needs to balance between exploring different strategies and exploiting currently optimal strategies within a given set of strategies. This is the underlying trade-off of a number of clinical neural engineering problems, including brain-computer interface, deep brain stimulation, and spinal cord injury therapy. In these systems, complex electronic and computational systems interact with the human central nervous system. A critical issue is how to control the agents to produce results which are optimal under some measure, for example, efficiently decoding the user's intention in a brain-computer interface or performs temporal and spatial specific stimulation in deep brain stimulation. This dissertation is motivated by electrical sipnal cord stimulation with high dimensional inputs(multi-electrode arrays). The stimulation is applied to promote the function and rehabilitation of the remaining neural circuitry below the spinal cord injury, and enable complex motor behaviors such as stepping and standing. To enable the careful tuning of these stimuli for each patient, the electrode arrays which deliver these stimuli have become increasingly more sophisticated, with a corresponding increase in the number of free parameters over which the stimuli need to be optimized. Since the number of stimuli is growing exponentially with the number of electrodes, algorithmic methods of selecting stimuli is necessary, particularly when the feedback is expensive to get.
\r\n\r\nIn many online learning settings, particularly those that involve human feedback, reliable feedback is often limited to pairwise preferences instead of real valued feedback. Examples include implicit or subjective feedback for information retrieval and recommender systems, such as clicks on search results, and subjective feedback on the quality of recommended care. Sometimes with real valued feedback, we require that the sampled function values exceed some prespecified ``safety'' threshold, a requirement that existing algorithms fail to meet. Examples include medical applications where the patients' comfort must be guaranteed; recommender systems aiming to avoid user dissatisfaction; and robotic control, where one seeks to avoid controls that cause physical harm to the platform.
\r\n\r\nThis dissertation provides online learning algorithms for several specific online decision-making problems. \\selfsparring optimizes the cumulative reward with relative feedback. RankComparison deals with ranking feedback. \\safeopt considers the optimization with real valued feedback and safety constraints. \\cduel is designed for specific spinal cord injury therapy.
\r\n\r\nA variant of \\cduel was implemented in closed-loop human experiments, controlling which epidural stimulating electrodes are used in the spinal cord of SCI patients. The results obtained are compared with concurrent stimulus tuning carried out by human experimenter. These experiments show that this algorithm is at least as effective as the human experimenter, suggesting that this algorithm can be applied to the more challenging problems of enabling and optimizing complex, sensory-dependent behaviors, such as stepping and standing in SCI patients.
\r\n\r\nIn order to get reliable quantitative measurements besides comparisons, the standing behaviors of paralyzed patients under spinal cord stimulation are evaluated. The potential of quantifying the quality of bipedal standing in an automatic approach is also shown in this work.
", "doi": "10.7907/Z9BK19DN", "publication_date": "2017", "thesis_type": "phd", "thesis_year": "2017" }, { "id": "thesis:10048", "collection": "thesis", "collection_id": "10048", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:02082017-062956314", "primary_object_url": { "basename": "Bruer-John-2017-thesis.pdf", "content": "final", "filesize": 7198711, "license": "other", "mime_type": "application/pdf", "url": "/10048/1/Bruer-John-2017-thesis.pdf", "version": "v4.0.0" }, "type": "thesis", "title": "Recovering Structured Low-rank Operators Using Nuclear Norms", "author": [ { "family_name": "Bruer", "given_name": "John Jacob", "orcid": "0000-0003-4590-3038", "clpid": "Bruer-John-Jacob" } ], "thesis_advisor": [ { "family_name": "Tropp", "given_name": "Joel A.", "clpid": "Tropp-J-A" } ], "thesis_committee": [ { "family_name": "Tropp", "given_name": "Joel A.", "clpid": "Tropp-J-A" }, { "family_name": "Hassibi", "given_name": "Babak", "clpid": "Hassibi-B" }, { "family_name": "Wierman", "given_name": "Adam C.", "clpid": "Wierman-A-C" }, { "family_name": "Yue", "given_name": "Yisong", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "This work considers the problem of recovering matrices and operators from limited and/or noisy observations. Whereas matrices result from summing tensor products of vectors, operators result from summing tensor products of matrices. These constructions lead to viewing both matrices and operators as the sum of \"simple\" rank-1 factors.
\r\n\r\nA popular line of work in this direction is low-rank matrix recovery, i.e., using linear measurements of a matrix to reconstruct it as the sum of few rank-1 factors. Rank minimization problems are hard in general, and a popular approach to avoid them is convex relaxation. Using the trace norm as a surrogate for rank, the low-rank matrix recovery problem becomes convex.
\r\n \r\nWhile the trace norm has received much attention in the literature, other convexifications are possible. This thesis focuses on the class of nuclear norms\u2014a class that includes the trace norm itself. Much as the trace norm is a convex surrogate for the matrix rank, other nuclear norms provide convex complexity measures for additional matrix structure. Namely, nuclear norms measure the structure of the factors used to construct the matrix.
\r\n\r\nTransitioning to the operator framework allows for novel uses of nuclear norms in recovering these structured matrices. In particular, this thesis shows how to lift structured matrix factorization problems to rank-1 operator recovery problems. This new viewpoint allows nuclear norms to measure richer types of structures present in matrix factorizations.
\r\n\r\nThis work also includes a Python software package to model and solve structured operator recovery problems. Systematic numerical experiments in operator denoising demonstrate the effectiveness of nuclear norms in recovering structured operators. In particular, choosing a specific nuclear norm that corresponds to the underlying factor structure of the operator improves the performance of the recovery procedures when compared, for instance, to the trace norm.\r\nApplications in hyperspectral imaging and self-calibration demonstrate the additional flexibility gained by utilizing operator (as opposed to matrix) factorization models.
", "doi": "10.7907/Z9F18WQS", "publication_date": "2017", "thesis_type": "phd", "thesis_year": "2017" }, { "id": "thesis:10218", "collection": "thesis", "collection_id": "10218", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:05292017-224523022", "primary_object_url": { "basename": "Hall_David_Thesis_2017.pdf", "content": "final", "filesize": 29724771, "license": "other", "mime_type": "application/pdf", "url": "/10218/1/Hall_David_Thesis_2017.pdf", "version": "v5.0.0" }, "type": "thesis", "title": "Advancing a Machine's Visual Awareness of People", "author": [ { "family_name": "Hall", "given_name": "David Christopher", "orcid": "0000-0003-3244-5744", "clpid": "Hall-David-Christopher" } ], "thesis_advisor": [ { "family_name": "Perona", "given_name": "Pietro", "clpid": "Perona-P" } ], "thesis_committee": [ { "family_name": "Perona", "given_name": "Pietro", "clpid": "Perona-P" }, { "family_name": "Abu-Mostafa", "given_name": "Yaser S.", "clpid": "Abu-Mostafa-Y-S" }, { "family_name": "Yue", "given_name": "Yisong", "clpid": "Yue-Yisong" }, { "family_name": "Kostina", "given_name": "Victoria", "clpid": "Kostina-V" }, { "family_name": "Munich", "given_name": "Mario E.", "clpid": "Munich-M-E" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Methods to advance a machine's visual awareness of people with a focus on understanding 'who is where' in video are presented. 'Who' is used in a broad sense that includes not only the identity of a person but attributes of that person as well. Efforts are focused on improving algorithms in four areas of visual recognition: detection, tracking, fine-grained classification and person reidentification.
\r\n\r\nEach of these problems appear to be quite different on the surface; however, there are two broader questions that are answered across each of the works. The first, the machine is able to make better predictions when it has access to the extra information that is available in video. The second, that it is possible to learn on-the-fly from single examples. How each work contributes to answering these over-arching questions as well as its specific contributions to the relevant problem domain are as follows:
\r\n\r\nThe first problem studied is one-shot, real-time, instance detection. Given a single image of a person, the task for the machine is to learn a detector that is specific to that individual rather than to an entire category such as faces or pedestrians. In subsequent images, the individual detector indicates the size and location of that particular person in the image. The learning must be done in real-time. To solve this problem, the proposed method starts with a pre-trained boosted category detector from which an individual-object detector is trained, with near-zero computational cost, through elementary manipulations of the thresholds of the category detector. Experiments on two challenging pedestrian and face datasets indicate that it is indeed possible to learn identity classifiers in real-time; besides being faster-trained, the proposed classifier has better detection rates than previous methods.
\r\n\r\nThe second problem studied is real-time tracking. Given the initial location of a target person, the task for the machine is to determine the size and location of the target person in subsequent video frames, in real-time. The method proposed for solving this problem treats tracking as a repeated detection problem where potential targets are identified with a pre-trained boosted person detector and identity across frames is established by individual-specific detectors. The individual-specific detectors are learnt using the method proposed to solve the first problem. The proposed algorithm runs in real-time and is robust to drift. The tracking algorithm is benchmarked against nine state-of-the-art trackers on two benchmark datasets. Results show that the proposed method is 10% more accurate and nearly as fast as the fastest of the competing algorithms, and it is as accurate but 20 times faster than the most accurate of the competing algorithms.
\r\n\r\nThe third problem studied is the fine-grained classification of people. Given an image of a person, the task for the machine is to estimate characteristics of that person such as age, clothing style, sex, occupation, social status, ethnicity, emotional state and/or body type. Since fine-grained classification using the entire human body is a relatively unexplored area, a large video dataset was collected. To solve this problem, a method that uses deep neural networks and video of a person is proposed. Results show that the class average accuracy when combining information from a sequence of images of an individual and then predicting the label is 3.5-7.1% better than independently predicting the label of each image, when severely under-represented classes are ignored.
\r\n\r\nThe final problem studied is person reidentification. Given an image of a person, the task for the machine is to find images that match the identity of that person from a large set of candidate images. This is a challenging task since images of the same individual can vary significantly due to changes in clothing, viewpoint, pose, lighting and background. The method proposed for solving this problem is a two-stage deep neural network architecture that uses body part patches as inputs rather than an entire image of a person. Experiments show that rank-1 matching rates increase by 22-25.6% on benchmark datasets when compared to state-of-the-art methods.
", "doi": "10.7907/Z9ST7MWC", "publication_date": "2017", "thesis_type": "phd", "thesis_year": "2017" }, { "id": "thesis:10292", "collection": "thesis", "collection_id": "10292", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:06052017-221842127", "type": "thesis", "title": "Boosting Boosting", "author": [ { "family_name": "Appel", "given_name": "Ron", "clpid": "Appel-Ron" } ], "thesis_advisor": [ { "family_name": "Perona", "given_name": "Pietro", "clpid": "Perona-P" } ], "thesis_committee": [ { "family_name": "Perona", "given_name": "Pietro", "clpid": "Perona-P" }, { "family_name": "Abu-Mostafa", "given_name": "Yaser S.", "clpid": "Abu-Mostafa-Y-S" }, { "family_name": "Bruck", "given_name": "Jehoshua", "clpid": "Bruck-J" }, { "family_name": "Hassibi", "given_name": "Babak", "clpid": "Hassibi-B" }, { "family_name": "Yue", "given_name": "Yisong", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "Machine learning is becoming prevalent in all aspects of our lives. For some applications, there is a need for simple but accurate white-box systems that are able to train efficiently and with little data.
\r\n\r\n\"Boosting\" is an intuitive method, combining many simple (possibly inaccurate) predictors to form a powerful, accurate classifier. Boosted classifiers are intuitive, easy to use, and exhibit the fastest speeds at test-time when implemented as a cascade. However, they have a few drawbacks: training decision trees is a relatively slow procedure, and from a theoretical standpoint, no simple unified framework for cost-sensitive multi-class boosting exists. Furthermore, (axis-aligned) decision trees may be inadequate in some situations, thereby stalling training; and even in cases where they are sufficiently useful, they don't capture the intrinsic nature of the data, as they tend to form boundaries that overfit.
\r\n\r\nMy thesis focuses on remedying these three drawbacks of boosting.\r\nCh.III outlines a method (called QuickBoost) that trains identical classifiers at an order of magnitude faster than before, based on a proof of a bound. In Ch.IV, a unified framework for cost-sensitive multi-class boosting (called REBEL) is proposed, both advancing theory and demonstrating empirical gains. Finally, Ch.V describes a novel family of weak learners (called Localized Similarities) that guarantee theoretical bounds and outperform decision trees and Neural Nets (as well as several other commonly used classification methods) on a range of datasets.
\r\n\r\nThe culmination of my work is an easy-to-use, fast-training, cost-sensitive multi-class boosting framework whose functionality is interpretable (since each weak learner is a simple comparison of similarity), and whose performance is better than Neural Networks and other competing methods. It is the tool that everyone should have in their toolbox and the first one they try.
", "doi": "10.7907/Z9N29V0J", "publication_date": "2017", "thesis_type": "phd", "thesis_year": "2017" }, { "id": "thesis:8888", "collection": "thesis", "collection_id": "8888", "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:05262015-094933189", "primary_object_url": { "basename": "CarlosGonzalez2015_thesis.pdf", "content": "final", "filesize": 1591603, "license": "other", "mime_type": "application/pdf", "url": "/8888/73/CarlosGonzalez2015_thesis.pdf", "version": "v9.0.0" }, "type": "thesis", "title": "Optimal Data Distributions in Machine Learning", "author": [ { "family_name": "Gonz\u00e1lez Palacios", "given_name": "Carlos Roberto", "clpid": "Gonz\u00e1lez-Palacios-Carlos-Roberto" } ], "thesis_advisor": [ { "family_name": "Abu-Mostafa", "given_name": "Yaser S.", "orcid": "0009-0005-1380-8345", "clpid": "Abu-Mostafa-Y-S" }, { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" } ], "thesis_committee": [ { "family_name": "Abu-Mostafa", "given_name": "Yaser S.", "orcid": "0009-0005-1380-8345", "clpid": "Abu-Mostafa-Y-S" }, { "family_name": "Perona", "given_name": "Pietro", "orcid": "0000-0002-7583-5809", "clpid": "Perona-P" }, { "family_name": "Vaidyanathan", "given_name": "P. P.", "orcid": "0000-0003-3003-7042", "clpid": "Vaidyanathan-P-P" }, { "family_name": "Bruck", "given_name": "Jehoshua", "orcid": "0000-0001-8474-0812", "clpid": "Bruck-J" }, { "family_name": "Yue", "given_name": "Yisong", "orcid": "0000-0001-9127-1989", "clpid": "Yue-Yisong" } ], "local_group": [ { "literal": "div_eng" } ], "abstract": "In the first part of the thesis we explore three fundamental questions that arise naturally when we conceive a machine learning scenario where the training and test distributions can differ. Contrary to conventional wisdom, we show that in fact mismatched training and test distribution can yield better out-of-sample performance. This optimal performance can be obtained by training with the dual distribution. This optimal training distribution depends on the test distribution set by the problem, but not on the target function that we want to learn. We show how to obtain this distribution in both discrete and continuous input spaces, as well as how to approximate it in a practical scenario. Benefits of using this distribution are exemplified in both synthetic and real data sets.
\r\n\r\nIn order to apply the dual distribution in the supervised learning scenario where the training data set is fixed, it is necessary to use weights to make the sample appear as if it came from the dual distribution. We explore the negative effect that weighting a sample can have. The theoretical decomposition of the use of weights regarding its effect on the out-of-sample error is easy to understand but not actionable in practice, as the quantities involved cannot be computed. Hence, we propose the Targeted Weighting algorithm that determines if, for a given set of weights, the out-of-sample performance will improve or not in a practical setting. This is necessary as the setting assumes there are no labeled points distributed according to the test distribution, only unlabeled samples.
\r\n\r\nFinally, we propose a new class of matching algorithms that can be used to match the training set to a desired distribution, such as the dual distribution (or the test distribution). These algorithms can be applied to very large datasets, and we show how they lead to improved performance in a large real dataset such as the Netflix dataset. Their computational complexity is the main reason for their advantage over previous algorithms proposed in the covariate shift literature.
\r\n\r\nIn the second part of the thesis we apply Machine Learning to the problem of behavior recognition. We develop a specific behavior classifier to study fly aggression, and we develop a system that allows analyzing behavior in videos of animals, with minimal supervision. The system, which we call CUBA (Caltech Unsupervised Behavior Analysis), allows detecting movemes, actions, and stories from time series describing the position of animals in videos. The method summarizes the data, as well as it provides biologists with a mathematical tool to test new hypotheses. Other benefits of CUBA include finding classifiers for specific behaviors without the need for annotation, as well as providing means to discriminate groups of animals, for example, according to their genetic line.
", "doi": "10.7907/Z9DR2SD5", "publication_date": "2015", "thesis_type": "phd", "thesis_year": "2015" } ]