-
Prospective Students
-
Faculty and Staff
-
Current Students
-
Alumni and Friends
Learning continuous functions parameterized by neural networks has become a novel paradigm for representing complex, high-dimensional data, offering many benefits like shift-invariance and resolution-independent representations. However, these models struggle with data that is discontinuous, noisy, non-linear, and ill-posed, largely due to their inability to capture diverse data characteristics in a unified manner. To overcome these challenges, we introduce Probabilistic Generative Neural Priors, a Bayesian-inspired regularization framework that integrates probabilistic generative models—such as Energy-based Models (EBMs), Score-based Diffusion Models (SBMs), and Variational Autoencoders (VAEs)—with task-specific neural networks like Neural Fields (NFs) and classification models. Our framework leverages generative models as probabilistic priors to provide essential information during inference network training, facilitating faster and more accurate predictions by directly utilizing the prior's outputs. We validate our approach through extensive experiments on a diverse set of applications, including non-linear physics-based partial differential equation (PDE) inverse problems, linear image inverse problems, physics-based topology optimization, and time-series classification. Our results show significant improvements in accuracy metrics, convergence speed, generalization and regularization performance compared to existing methods, across all considered applications.
Busyness, or how busy someone is, has increasingly become a topic of conversation in day-to-day life. Research has previously explored how people use their time and how people perceive their available time, or lack thereof, but there is no clear answer as to why people tell others that they are busy and what it is they are trying to accomplish by doing so. Drawing on impression management research, this paper proposes that people signal to others that they are busy so that the audience has a positive impression of them. The concept of the busyness facade is introduced, which includes behaviors and verbal statements that are intentionally enacted by individuals to signal to others that they have a lot to do or limited available time. Exactly how and why people engage in this busyness facade is explored in two studies using semi-structured interviews and an online, vignette survey. Overall, evidence is found for the existence of busyness facades and a better understanding of how people display busyness is gained, but the studies are unable to identify a clear motive for why busyness facades would be used as an impression management tactic. Additional findings and research directions are discussed.
Freeform optics offer improved optical systems, but their complex shapes challenge traditional measurement methods. Cost-effective solutions are needed, especially for applications where expensive methods are impractical. Non-interferometric methods are a good alternative, but their accuracy can be limited. This dissertation aims to develop an accessible calibration method that improves the accuracy of these methods and enables the measurement of both refractive and reflective elements. The results are presented in three articles. The first article focuses on the calibration method and a new metrology approach that directly measures ray deflections, simplifying the process. The second article analyzes a new technique for converting wavefront data to height information and proposes a calibration process to improve accuracy. The third article tackles the issue of parasitic reflections by using a data-driven approach. This work significantly advances ray trace-based optical metrology and has numerous applications, particularly in the measurement and alignment of freeform optics.
Causal modeling provides us with powerful counterfactual reasoning and interventional mechanisms to generate predictions under various what-if scenarios. Nevertheless, uncovering causal relationships from observational data presents a considerable challenge, as unobserved confounders, limited sample sizes, and variations in distributions can give rise to misleading cause-effect associations. Models relying on these relationships may perform poorly when spurious correlations do not hold in test cases. To mitigate these challenges, researchers augment causal learning with known causal relations. This dissertation first investigates the incorporation of domain knowledge in structure learning by introducing additional constraints that convey qualitative knowledge about causal relationships. The experimental designs are specifically equipped to evaluate the role of domain knowledge. Secondly, a concept-driven approach is implemented to determine the advantages of incorporating concept-level prior knowledge. Given the invariant nature of causal relationships, the study then showcases the broader applicability of incorporating domain knowledge by employing a machine learning method for learning adsorption energies, illustrating the advantages of harnessing domain knowledge to obtain invariant molecular representations in catalyst screening. Finally, a novel approach is introduced to enhance robustness and out-of-distribution generalization by leveraging gradient agreement across different environments to identify reliable features. Collectively, these experimental designs advance causal discovery and robust machine learning by utilizing prior knowledge and relational invariances, paving the way for future research on integrating domain knowledge and invariance principles into the learning process.
Machine learning has risen to the forefront of scientific research due to its unparalleled predictive capabilities. As a result, researchers have become increasingly interested in uncovering the underlying causal structures that govern the relationships between variables in a system. These causal structures, often represented as directed acyclic graphs (DAGs), provide insights into how changes in one variable may directly or indirectly affect other variables, enabling a deeper understanding of the complex interactions within the system. While it is essential to constrain a model by minimizing spurious correlations and conducting "What-If" analyses, learning causal relationships from observational data, known as causal discovery, remains an active and challenging research area. This is due to factors like finite sampling, unobserved confounding factors, and measurement errors. Current approaches, including constraint-based and score-based methods, often struggle with high computational complexity because of the combinatorial nature of estimating DAGs. Inspired by the workshop on the Causality Challenge 'Cause-Effect Pair' at the Neural Information Processing Systems in 2013, this dissertation adopts a novel approach, generating a probability distribution over all possible graphs based on cause-effect pair features proposed in response to the workshop challenge.
The primary goal of this study is to develop new methods that leverage this probabilistic information and assess their performance. Furthermore, this work introduces a novel causal feature selection (CFS) algorithm using this approach and the establishment of a new evaluation criterion for CFS. To further enhance experimental performance, this dissertation proposes the use of a Graph Neural Networks (GNNs)--based probabilistic predictive framework for causal discovery. Conventional causal discovery algorithms face significant challenges in dealing with large-scale observational datasets and capturing global structural information. The GNN-based approach addresses these limitations, enabling the learning of complex causal structures directly from data augmented with statistical and information-theoretic measures. The proposed framework represents a significant leap forward in causal discovery, offering improved accuracy and scalability in both synthetic and real-world datasets, as well as introducing a novel synergy between probabilistic learning and causal graph analysis.
In addition to the methodological advancements, this dissertation includes an application of counterfactual analysis to study affective polarization on social media. By comparing scenarios with and without specific influencer-led conversations on platforms like Twitter, I analyze the impact of these conversations on public sentiment. This application highlights the practical implications of the proposed causal modeling techniques, demonstrating their utility in understanding real-world issues and contributing to the broader field of social media analysis.
Electrical fires caused by arc faults necessitate advanced detection methods for improved safety and reliability in AC and DC power systems. This research introduces innovative artificial intelligence (AI)-based methods for the efficient detection of arc faults, enhancing both safety and reliability in electrical systems. For AC systems, we developed a convolutional neural network (CNN)-based arc fault detection algorithm that autonomously extracts arc fault features without manual thresholding. Using raw current as input, the algorithm achieves an arc fault detection accuracy of 99.47%. Additionally, this research determined an optimal sampling rate of 10 kHz for the input current. The model's efficacy was verified using the Raspberry Pi 3B platform.
While traditional CNN algorithms have high accuracy, they require optimization for real-time arc fault detection on resource-limited hardware. To address this, we proposed a lightweight CNN architecture combined with a model compression technique using a knowledge distillation-based teacher-student algorithm. This model maintains a high detection accuracy of 99.31% and operates with an impressively minimal runtime of 0.20 milliseconds per sample when implemented on the Raspberry Pi 3B platform. This performance demonstrates its suitability for commercial embedded microcontrollers (MCUs) with limited computational capability.
Extending our research to DC systems, we introduced a cost-effective, AI-driven Arc Fault Circuit Interrupter (AFCI) for DC applications. Utilizing an STM32 MCU and a silicon carbide (SiC) MOSFET-based solid-state circuit breaker (SSCB), the proposed method achieves a detection accuracy of 98.15% with a remarkable arc fault interruption time of 25 milliseconds. This AFCI solution stands out for its rapid response and high reliability, promising significant improvements in safety for DC systems.
Together, these contributions signify a leap forward in electrical safety, presenting viable solutions for the timely detection and interruption of arc faults in AC and DC systems. The outcomes of this research are expected to influence future standards and practices in electrical safety management across residential, commercial, and industrial sectors.
Social media has emerged as a common platform for knowledge sharing and exchange in online communities. However, it has also become a hotbed for the diffusion of irregular content. Content moderation is crucial for maintaining a safe and healthy online environment by regulating the distribution of user-generated content (UGC).
Engaging users in content moderation fosters a sense of shared responsibility and empowers them to actively shape the environment of online communities. Leveraging the expertise of moderators leads to a deeper contextual understanding of content, thereby improving the overall consistency and legitimacy of content moderation in compliance with community or platform guidelines. Nevertheless, the collaborative effort of a more inclusive and community-driven moderation process remains unexplored by previous studies. While there is increasing attention to fairness, transparency, and ethics in content moderation, prior research often assesses content moderation perceptions of users, platforms, moderators, and bystanders in isolation. This results in a lack of comprehensive understanding of user perceptions in content moderation decision-making.
To address these limitations, this research proposes UMCollab, a user-moderator collaborative content moderation framework that incorporates the dynamics of user engagement and the domain knowledge of moderators into deep learning models to facilitate content moderation decision-making. Additionally, this research empirically investigates user perceptions of content moderation from the perspectives of content familiarity, content diversity, and user roles.
UMCollab leverages graph learning to model user engagement, which is further enhanced by the credibility and stance of users' online discussions. It also employs attention mechanisms to learn the domain knowledge of moderators based on their decisions regarding UGC per online community rules. Moreover, this study conducts an online user study by asking participants with diverse online engagement backgrounds and roles to complete a series of content moderation decision-making tasks and evaluate their perceptions of content moderation.
The findings of this dissertation research hold significant promise for promoting effectiveness, fairness, transparency, and community ownership in moderating UGC in social media, offering opportunities to improve the safety and success of online communities.
The advancement of medical imaging has significantly enhanced the ability to diagnose, monitor, and treat cancer. This dissertation focuses on the development of deep learning methodologies for the segmentation and registration of medical images, specifically Positron Emission Tomography (PET), Computed Tomography (CT), Magnetic Resonance Imaging (MRI), and pathology images, to improve the accuracy and efficiency of cancer diagnosis and treatment planning.
Segmentation, the process of delineating anatomical structures and pathological regions, is a crucial step in medical image analysis. This work introduces novel high-precision deep learning models for the automatic segmentation of tumors and organs at risk (OARs). These models utilize convolutional neural networks (CNNs) and transformer-based architectures to handle the complexities and variations inherent in PET, CT, and MRI. The segmentation models are trained on multi-modal imaging datasets, incorporating advanced techniques such as data augmentation, transfer learning, and ensemble learning to enhance robustness and generalization. Evaluation on various datasets demonstrates that these models achieve superior performance compared to traditional methods, with significant improvements in accuracy and reliability.
Registration, which aligns images from different modalities or time points, is another critical component in the analysis of medical images. This dissertation presents advanced deep learning approaches for the registration of CT, MRI, and pathology images, leveraging deep neural networks (DNNs) and unsupervised learning techniques. The proposed registration methods employ spatial transformer networks (STNs) and other novel architectures to learn complex spatial transformations directly from the data, enabling accurate alignment of multi-modal images. These approaches are designed to be computationally efficient and scalable, facilitating their integration into clinical workflows.
Our final goal is to streamline these deep learning methods to real clinical applications. This dissertation explores the practical applications of the developed models, including their deployment in microservices for common radiotherapy imaging tasks. The models are made accessible via Python scripts for clinical treatment planning software such as RayStation, allowing seamless integration into existing clinical systems. Evaluation using images and treatment planning data for prostate cancer underscores the potential of these models to enhance the quality of treatment planning and streamline the overall process of planning, response assessment, and adaptation. Additionally, this dissertation investigates the potential of federated learning for collaborative model training across multiple institutions without sharing sensitive patient data. This approach could enhance model robustness and generalizability by leveraging diverse datasets from various sources.
In conclusion, this dissertation explores the critical component of medical imaging for cancer diagnosis, monitoring, and treatment with advanced deep learning methods. We hope these innovative techniques developed in this research pave the way for more precise, efficient, and individualized patient care in oncology.
People adopt security technologies and make security decisions based on their perceptions, or mental models, of what risks they have and what they can do to protect their devices. Thus, people rely on their mental models to decide how to use their computing devices and the consequences of these actions. Understanding why users make security decisions and addressing the misconceptions in their mental models, specifically regarding security risks, can help prevent security mistakes made by users and help determine how to help users make good security decisions. This dissertation seeks to understand how users perceive security risks, why they make security-related decisions, and where they have misconceptions. In my dissertation, I examine how users' mental models of security and privacy differ by device platform, how that impacts how people use and interact with applications on each platform, and how user’s mental models can be used to influence adoption of good device security practices. I will present the results of three user studies exploring user mental models of security and privacy and how users need an increasing awareness of security risks and measures across all types of computing platforms in order to adopt appropriate practices to protect themselves and their information.
Physical activity offers a range of health benefits but can be difficult to initiate and maintain. Self-regulatory processes are one route to understanding health behavior change. While affective mechanisms like positive affect and reward processing provide a valuable neurobiological pathway to elucidate individual motivations for physical activity. Theories and models that emphasize the role of affect and intrinsic and extrinsic motivations are applied in the current study to deepen our understanding of physical activity behaviors. Psychosocial sources of individual motivations, such as improving mood or changing one’s physical appearance, can provide insight into ways to alter affective-cognitive mechanisms to encourage sustainable physical activity behaviors. The present study applied a mixed-method approach to elucidate themes of affect and motivation in social media content. A sample of 2,585 Twitter posts were collected in mid-July 2022. A motivational and health behavior change qualitative codebook was first developed to guide thematic coding analyses. Thematic coding results revealed a high frequency of extrinsic motivation and goal and change-oriented facilitators of physical activity. Intrinsic motivation included the highest percentage of positive attitudes compared to other motivation types. Health-oriented themes, satisfaction, dissatisfaction with physical appearance, and weight loss were also relevant. LIWC-22 analyses supported the role of positive affect and informed health themes. BERT topic modeling analyses provided overarching physical activity topic themes for motivation and physical activity. Interpretations of the current results were presented, and future directions were suggested.
