Artificial Intelligence and Robotics Commons^™

Policy Gradient Methods: Analysis, Misconceptions, And Improvements, Christopher P. Nota

Doctoral Dissertations

Policy gradient methods are a class of reinforcement learning algorithms that optimize a parametric policy by maximizing an objective function that directly measures the performance of the policy. Despite being used in many high-profile applications of reinforcement learning, the conventional use of policy gradient methods in practice deviates from existing theory. This thesis presents a comprehensive mathematical analysis of policy gradient methods, uncovering misconceptions and suggesting novel solutions to improve their performance. We first demonstrate that the update rule used by most policy gradient methods does not correspond to the gradient of any objective function due to the way the …

Multi-Slam Systems For Fault-Tolerant Simultaneous Localization And Mapping, Samer Nashed

Doctoral Dissertations

Mobile robots need accurate, high fidelity models of their operating environments in order to complete their tasks safely and efficiently. Generating these models is most often done via Simultaneous Localization and Mapping (SLAM), a paradigm where the robot alternatively estimates the most up-to-date model of the environment and its position relative to this model as it acquires new information from its sensors over time. Because robots operate in many different environments with different compute, memory, sensing, and form constraints, the nature and quality of information available to individual instances of different SLAM systems varies substantially. `One-size-fits-all' solutions are thus exceedingly …

Generalized Differentiable Neural Architecture Search With Performance And Stability Improvements, Emily J. Herron

Doctoral Dissertations

This work introduces improvements to the stability and generalizability of Cyclic DARTS (CDARTS). CDARTS is a Differentiable Architecture Search (DARTS)-based approach to neural architecture search (NAS) that uses a cyclic feedback mechanism to train search and evaluation networks concurrently, thereby optimizing the search process by enforcing that the networks produce similar outputs. However, the dissimilarity between the loss functions used by the evaluation networks during the search and retraining phases results in a search-phase evaluation network, a sub-optimal proxy for the final evaluation network utilized during retraining. ICDARTS, a revised algorithm that reformulates the search phase loss functions to ensure …

Exact Models, Heuristics, And Supervised Learning Approaches For Vehicle Routing Problems, Zefeng Lyu

Doctoral Dissertations

This dissertation presents contributions to the field of vehicle routing problems by utilizing exact methods, heuristic approaches, and the integration of machine learning with traditional algorithms. The research is organized into three main chapters, each dedicated to a specific routing problem and a unique methodology. The first chapter addresses the Pickup and Delivery Problem with Transshipments and Time Windows, a variant that permits product transfers between vehicles to enhance logistics flexibility and reduce costs. To solve this problem, we propose an efficient mixed-integer linear programming model that has been shown to outperform existing ones. The second chapter discusses a practical …

Generative Adversarial Game With Tailored Quantum Feature Maps For Enhanced Classification, Anais Sandra Nguemto Guiawa

Doctoral Dissertations

In the burgeoning field of quantum machine learning, the fusion of quantum computing and machine learning methodologies has sparked immense interest, particularly with the emergence of noisy intermediate-scale quantum (NISQ) devices. These devices hold the promise of achieving quantum advantage, but they grapple with limitations like constrained qubit counts, limited connectivity, operational noise, and a restricted set of operations. These challenges necessitate a strategic and deliberate approach to crafting effective quantum machine learning algorithms.

This dissertation revolves around an exploration of these challenges, presenting innovative strategies that tailor quantum algorithms and processes to seamlessly integrate with commercial quantum platforms. A …

Towards Robust Long-Form Text Generation Systems, Kalpesh Krishna

Doctoral Dissertations

Text generation is an important emerging AI technology that has seen significant research advances in recent years. Due to its closeness to how humans communicate, mastering text generation technology can unlock several important applications such as intelligent chat-bots, creative writing assistance, or newer applications like task-agnostic few-shot learning. Most recently, the rapid scaling of large language models (LLMs) has resulted in systems like ChatGPT, capable of generating fluent, coherent and human-like text. However, despite their remarkable capabilities, LLMs still suffer from several limitations, particularly when generating long-form text. In particular, (1) long-form generated text is filled with factual inconsistencies to …

Quantifying And Enhancing The Security Of Federated Learning, Virat Vishnu Shejwalkar

Doctoral Dissertations

Federated learning is an emerging distributed learning paradigm that allows multiple users to collaboratively train a joint machine learning model without having to share their private data with any third party. Due to many of its attractive properties, federated learning has received significant attention from academia as well as industry and now powers major applications, e.g., Google's Gboard and Assistant, Apple's Siri, Owkin's health diagnostics, etc. However, federated learning is yet to see widespread adoption due to a number of challenges. One such challenge is its susceptibility to poisoning by malicious users who aim to manipulate the joint machine learning …

Learning To See With Minimal Human Supervision, Zezhou Cheng

Doctoral Dissertations

Deep learning has significantly advanced computer vision in the past decade, paving the way for practical applications such as facial recognition and autonomous driving. However, current techniques depend heavily on human supervision, limiting their broader deployment. This dissertation tackles this problem by introducing algorithms and theories to minimize human supervision in three key areas: data, annotations, and neural network architectures, in the context of various visual understanding tasks such as object detection, image restoration, and 3D generation. First, we present self-supervised learning algorithms to handle in-the-wild images and videos that traditionally require time-consuming manual curation and labeling. We demonstrate that …

Foundations Of Node Representation Learning, Sudhanshu Chanpuriya

Doctoral Dissertations

Low-dimensional node representations, also called node embeddings, are a cornerstone in the modeling and analysis of complex networks. In recent years, advances in deep learning have spurred development of novel neural network-inspired methods for learning node representations which have largely surpassed classical 'spectral' embeddings in performance. Yet little work asks the central questions of this thesis: Why do these novel deep methods outperform their classical predecessors, and what are their limitations? We pursue several paths to answering these questions. To further our understanding of deep embedding methods, we explore their relationship with spectral methods, which are better understood, and show …

Bayesian Structural Causal Inference With Probabilistic Programming, Sam A. Witty

Doctoral Dissertations

Reasoning about causal relationships is central to the human experience. This evokes a natural question in our pursuit of human-like artificial intelligence: how might we imbue intelligent systems with similar causal reasoning capabilities? Better yet, how might we imbue intelligent systems with the ability to learn cause and effect relationships from observation and experimentation? Unfortunately, reasoning about cause and effect requires more than just data: it also requires partial knowledge about data generating mechanisms. Given this need, our task then as computational scientists is to design data structures for representing partial causal knowledge, and algorithms for updating that knowledge in …

Effective And Efficient Transfer Learning In The Era Of Large Language Models, Tu Vu

Doctoral Dissertations

Substantial progress has been made in the field of natural language processing (NLP) due to the advent of large language models (LLMs)—deep neural networks with millions or billions of parameters pre-trained on large amounts of unlabeled data. However, these models have common weaknesses, including degenerate performance in data-scarce scenarios, and substantial computational resource requirements. This thesis aims to develop methods to address these limitations for improved applicability and performance of LLMs in resource-constrained settings with limited data and/or computational resources. To address the need for labeled data in data-scarce scenarios, I present two methods, in Chapter 2 and Chapter 3, …

Graph Representation Learning With Box Embeddings, Dongxu Zhang

Doctoral Dissertations

Graphs are ubiquitous data structures, present in many machine-learning tasks, such as link prediction of products and node classification of scientific papers. As gradient descent drives the training of most modern machine learning architectures, the ability to encode graph-structured data using a differentiable representation is essential to make use of this data. Most approaches encode graph structure in Euclidean space, however, it is non-trivial to model directed edges. The naive solution is to represent each node using a separate "source" and "target" vector, however, this can decouple the representation, making it harder for the model to capture information within longer …

An Introspective Approach For Competence-Aware Autonomy, Connor Basich

Doctoral Dissertations

Building and deploying autonomous systems in the open world has long been a goal of both the artificial intelligence (AI) and robotics communities. From autonomous driving, to health care, to office assistance, these systems have the potential to transform society and alter our everyday lives. The open world, however, presents numerous challenges that question the typical assumptions made by the models and frameworks often used in contemporary AI and robotics. Systems in the open world are faced with an unconstrained and non-stationary environment with a range of heterogeneous actors that is too complex to be modeled in its entirety. Moreover, …

Optimizing Collective Communication For Scalable Scientific Computing And Deep Learning, Jiali Li

Doctoral Dissertations

In the realm of distributed computing, collective operations involve coordinated communication and synchronization among multiple processing units, enabling efficient data exchange and collaboration. Scientific applications, such as simulations, computational fluid dynamics, and scalable deep learning, require complex computations that can be parallelized across multiple nodes in a distributed system. These applications often involve data-dependent communication patterns, where collective operations are critical for achieving high performance in data exchange. Optimizing collective operations for scientific applications and deep learning involves improving the algorithms, communication patterns, and data distribution strategies to minimize communication overhead and maximize computational efficiency.

Within the context of this …

Rigorous Experimentation For Reinforcement Learning, Scott M. Jordan

Doctoral Dissertations

Scientific fields make advancements by leveraging the knowledge created by others to push the boundary of understanding. The primary tool in many fields for generating knowledge is empirical experimentation. Although common, generating accurate knowledge from empirical experiments is often challenging due to inherent randomness in execution and confounding variables that can obscure the correct interpretation of the results. As such, researchers must hold themselves and others to a high degree of rigor when designing experiments. Unfortunately, most reinforcement learning (RL) experiments lack this rigor, making the knowledge generated from experiments dubious. This dissertation proposes methods to address central issues in …

Learning From Sequential User Data: Models And Sample-Efficient Algorithms, Aritra Ghosh

Doctoral Dissertations

Recent advances in deep learning have made learning representation from ever-growing datasets possible in the domain of vision, natural language processing (NLP), and robotics, among others. However, deep networks are notoriously data-hungry; for example, training language models with attention mechanisms sometimes requires trillions of parameters and tokens. In contrast, we can often access a limited number of samples in many tasks. It is crucial to learn models from these `limited' datasets. Learning with limited datasets can take several forms. In this thesis, we study how to select data samples sequentially such that downstream task performance is maximized. Moreover, we study …

Constrained Collective Movement In Human-Robot Teams, Joshua Fagan

Doctoral Dissertations

This research focuses on improving human-robot co-navigation for teams of robots and humans navigating together as a unit while accomplishing a desired task. Frequently, the team’s co-navigation is strongly influenced by a predefined Standard Operating Procedure (SOP), which acts as a high-level guide for where agents should go and what they should do. In this work, I introduce the concept of Constrained Collective Movement (CCM) of a team to describe how members of the team perform inter-team and intra-team navigation to execute a joint task while balancing environmental and application-specific constraints. This work advances robots’ abilities to participate along side …

Labeled Modules In Programs That Evolve, Anil K. Saini

Doctoral Dissertations

Multiple methods have been developed for Inductive Program Synthesis, i.e., synthesizing programs consistent with a set of input-output examples. One such method is genetic programming, which searches for programs with desirable properties from the space of all possible programs through an iterated process of variation and selection that is inspired by natural evolution. Genetic programming has been successful in solving problems from multiple domains. These problems are often challenging because of the range of data types and control structures they require to be solved. Nonetheless, there are many programming problems that are routinely solved by human programmers that cannot be …

Low Resource Language Understanding In Voice Assistants, Subendhu Rongali

Doctoral Dissertations

Voice assistants such as Amazon Alexa, Apple Siri, and Google Assistant have become ubiquitous. They rely on spoken language understanding, which typically consists of an Automatic Speech Recognition (ASR) component and a Natural Language Understanding (NLU) component. ASR takes user speech as input and generates a text transcription. NLU takes the text transcription as input and generates a semantic parse to identify the requested actions, called intents (play music, turn on lights, etc.) and any relevant entities, called slots (which song to play? which lights to turn on?).

These components require massive amounts of training data to achieve good performance. …

Neural Approaches For Language-Agnostic Search And Recommendation, Hamed Rezanejad Asl Bonab

Doctoral Dissertations

There are significant efforts toward developing better neural approaches for information retrieval problems. However, the vast majority of these studies are conducted using English-only data. In fact, trends and statistics of non-English content and users on the Internet show exponential growth and that novel information retrieval systems need to be language-agnostic; they need to bridge the language barrier between users and content, leverage data from high-resource settings for lower-resourced settings, and be able to extend to new languages and local markets easily. To this end, we focus on search and recommendation as two vital components of information systems. We explore …

Answer Similarity Grouping And Diversification In Question Answering Systems, Lakshmi Nair Vikraman

Doctoral Dissertations

The rise in popularity of mobile and voice search has led to a shift in IR from document to passage retrieval for non-factoid questions. Various datasets such as MSMarco, as well as efficient retrieval models have been developed to identify single best answer passages for this task. However, such models do not specifically address questions which could have multiple or alternative answers. In this dissertation, we focus on this new research area that involves studying answer passage relationships and how this could be applied to passage retrieval tasks. We first create a high quality dataset for the answer passage similarity …

Approximate Bayesian Deep Learning For Resource-Constrained Environments, Meet Prakash Vadera

Doctoral Dissertations

Deep learning models have shown promising results in areas including computer vision, natural language processing, speech recognition, and more. However, existing point estimation-based training methods for these models may result in predictive uncertainties that are not well calibrated, including the occurrence of confident errors. Approximate Bayesian inference methods can help address these issues in a principled way by accounting for uncertainty in model parameters. However, these methods are computationally expensive both when computing approximations to the parameter posterior and when using an approximate parameter posterior to make predictions. They can also require significantly more storage than point-estimated models. In this …

Controllable Neural Synthesis For Natural Images And Vector Art, Difan Liu

Doctoral Dissertations

Neural image synthesis approaches have become increasingly popular over the last years due to their ability to generate photorealistic images useful for several applications, such as digital entertainment, mixed reality, synthetic dataset creation, computer art, to name a few. Despite the progress over the last years, current approaches lack two important aspects: (a) they often fail to capture long-range interactions in the image, and as a result, they fail to generate scenes with complex dependencies between their different objects or parts. (b) they often ignore the underlying 3D geometry of the shape/scene in the image, and as a result, they …

Probabilistic Commonsense Knowledge, Xiang Li

Doctoral Dissertations

Commonsense knowledge is critical to achieving artificial general intelligence. This shared common background knowledge is implicit in all human communication, facilitating efficient information exchange and understanding. But commonsense research is hampered by its immense quantity of knowledge because an explicit categorization is impossible. Furthermore, a plumber could repair a sink in a kitchen or a bathroom, indicating that common sense reveals a probable assumption rather than a definitive answer. To align with these properties of commonsense fundamentally, we want to not only model but also evaluate such knowledge human-like using abstractions and probabilistic principles. Traditional combinatorial probabilistic models, e.g., probabilistic …

Modeling The Multi-Mode Distribution In Self-Supervised Language Models, Haw-Shiuan Chang

Doctoral Dissertations

Self-supervised large language models (LMs) have become a highly-influential and foundational tool for many NLP models. For this reason, their expressivity is an important topic of study. In near-universal practice, given the language context, the model predicts a word from the vocabulary using a single embedded vector representation of both context and dictionary entries. Note that the context sometimes implies that the distribution over predicted words should be multi-modal in embedded space. However, the context’s single-vector representation provably fails to capture such a distribution. To address this limitation, we propose to represent context with multiple vector embeddings, which we term …

Combinatorial Algorithms For Graph Discovery And Experimental Design, Raghavendra K. Addanki

Doctoral Dissertations

In this thesis, we study the design and analysis of algorithms for discovering the structure and properties of an unknown graph, with applications in two different domains: causal inference and sublinear graph algorithms. In both these domains, graph discovery is possible using restricted forms of experiments, and our objective is to design low-cost experiments. First, we describe efficient experimental approaches to the causal discovery problem, which in its simplest form, asks us to identify the causal relations (edges of the unknown graph) between variables (vertices of the unknown graph) of a given system. For causal discovery, we study algorithms …

Better Understanding Genomic Architecture With The Use Of Applied Statistics And Explainable Artificial Intelligence, Jonathon C. Romero

Doctoral Dissertations

With the continuous improvements in biological data collection, new techniques are needed to better understand the complex relationships in genomic and other biological data sets. Explainable Artificial Intelligence (X-AI) techniques like Iterative Random Forest (iRF) excel at finding interactions within data, such as genomic epistasis. Here, the introduction of new methods to mine for these complex interactions is shown in a variety of scenarios. The application of iRF as a method for Genomic Wide Epistasis Studies shows that the method is robust in finding interacting sets of features in synthetic data, without requiring the exponentially increasing computation time of many …

Nonparametric Contextual Reasoning For Question Answering Over Large Knowledge Bases, Rajarshi Das

Doctoral Dissertations

Question answering (QA) over knowledge bases provides a user-friendly way of accessing the massive amount of information stored in them. We have experienced tremendous progress in the performance of QA systems, thanks to the recent advancements in representation learning by deep neural models. However, such deep models function as black boxes with an opaque reasoning process, are brittle, and offer very limited control (e.g. for debugging an erroneous model prediction). It is also unclear how to reliably add or update knowledge stored in their model parameters. This thesis proposes nonparametric models for question answering that disentangle logic from knowledge. For …

Iterative Random Forest Based High Performance Computing Methods Applied To Biological Systems And Human Health, Angelica M. Walker

Doctoral Dissertations

As technology improves, the field of biology has increasingly utilized high performance computing techniques to analyze big data and provide insights into biological systems. A reproducible, efficient, and effective method is required to analyze these large datasets of varying types into interpretable results. Iterative Random Forest (iRF) is an explainable supervised learner that makes few assumptions about the relationships between variables and is able to capture complex interactions that are common in biological systems. This forest based learner is the basis of iRF-Leave One Out Prediction (iRF-LOOP), an algorithm that uses a matrix of data to produce all-to-all predictive networks. …

Toward Scalable Morphogenetic Engineering: Natural Computing In Sph Swarm Control, Allen C. Mcbride

Doctoral Dissertations

Artificial morphogenesis (or morphogenetic engineering) seeks inspiration from developmental biology to engineer self-organizing systems. The Morphgen language uses partial differential equations (PDEs) to express artificial morphogenetic processes as spatial fields describing large numbers of agents in the continuum limit. I present an approach to compile such systems of PDEs by discretizing their behavior to derive controllers for finite numbers of agents of finite size. This approach builds on a generalization of methods to control swarms of robots based on the computational fluid dynamics technique of smoothed particle hydrodynamics (SPH). I address potential scalability and efficiency challenges in SPH robotics by …