Subscribe to RSS Feed| Start Date: 22 Nov 2024; Start Time: 11:30AM - 01:00PM Title: Robots with Dynamics: Efficient Motion Planning and Analysis of Controllers via Machine Learning Bio: Speaker: Abstract: This thesis aims to improve the efficiency and robustness of motion planning for robots with significant dynamics, leveraging both advances in machine learning as well as contributions in algorithmic and foundational techniques. The key objectives are to (a) efficiently compute safe open-loop trajectories that obey non-trivial robot dynamics so that they are easy to follow with closed-loop controllers, and (b) efficiently analyze and characterize the capabilities of closed-loop robot controllers to enable safe real-world deployment.This effort starts by exploring alternatives to the standard methodology of generating control sequences in sampling-based planning for systems with dynamics. Typically, these methods rely on random controls, which are useful to argue desirable properties, but which lead to slow convergence and low-quality solutions in practice.To address this, the thesis first proposes using machine learning to train goal-reaching controllers via reinforcement learning. Such learned controllers can be integrated with sampling-based planners and help guide the expansion of the underlying planning structure towards the global goal. This is shown to lead to the faster discovery of high-quality trajectories on mobile robot navigation problems, including for physically-simulated challenges with uneven terrains.In addition, this thesis proposes the offline construction of a “roadmap with gaps” data structure for systems with dynamics, which can express the learned controller's reachability capabilities in a target environment. Online, the sampling-based planner uses the “roadmap with gaps” to promote the fact discovery of high-quality trajectories to the goal. The overall approach enhances the efficiency of motion planning in various benchmarks, including physics-based simulations of vehicular systems and aerial robots.The open-loop solutions generated by sampling-based planners require closed-loop feedback control for reliable real-world execution. To this end, the thesis first integrates techniques for identifying approximate analytical models of the robot's dynamics that allow fast motion planning and reduce the model gap. It then focuses on achieving closed-loop operation at both the planning and control levels by proposing a safe replanning framework for kinodynamic motion planning and integrating feedback controllers that reason about robot dynamics. These contributions allow for safe and efficient tracking of planned trajectories on a physical platform.Concurrently, the thesis also addresses the challenge of understanding the global dynamics of robot controllers, including learned ones, which is crucial for safe deployment of such solutions and the composition of controllers. A topological framework (Morse Graphs) is leveraged, and data-driven modeling approaches are proposed to enable data-efficient characterization of controller attractors and their regions of attraction, even for high-dimensional systems.Finally, the thesis contributes an open-source software library, which provides a flexible and efficient framework for integrating machine learning methods into kinodynamic planning and control. Location: SPR-402, 1 Spring St, New Brunswick Committee: Professor Kostas Bekris (Chair) Associate Professor Abdeslam Boularias Associate Professor Jingjin Yu Associate Professor Joydeep Biswas (external -- UT Austin). |
| Start Date: 22 Nov 2024; Start Time: 02:00PM - 03:30PM Title: Learning to Reason with LLMs Bio: Noam Brown is a research scientist at OpenAI investigating reasoning and multi-agent AI. He co-created Libratus and Pluribus, the first AIs to defeat top humans in two-player no-limit poker and multiplayer no-limit poker, respectively, and Cicero, the first AI to achieve human-level performance in the natural language strategy game Diplomacy. He has received the Marvin Minsky Medal for Outstanding Achievements in AI, was named one of MIT Tech Review's 35 Innovators Under 35, and his work on Pluribus was named by Science as one of the top 10 scientific breakthroughs of 2019. Noam received his PhD from Carnegie Mellon University and his BA from Rutgers University. Speaker: Abstract: Large language models (LLMs) have demonstrated remarkable capabilities in generating coherent text and completing various natural language tasks. Nevertheless, their ability to perform complex, general reasoning has remained limited. In this talk, I will describe OpenAI's new o1 model, an LLM trained via reinforcement learning to generate a hidden chain of thought before its response. We have found that the performance of o1 consistently improves with more reinforcement learning compute and with more inference compute. o1 surpasses previous state-of-the-art models in a variety of benchmarks that require reasoning, including mathematics competitions, programming contests, and advanced science question sets. I will discuss the implications of scaling this paradigm even further. Location: CoRE 301 Committee: |
| Start Date: 26 Nov 2024; Start Time: 09:00AM - 10:15AM Title: Neural Unsupervised Structure-Aware Representation Learning Bio: Speaker: Abstract: Although deep learning models have shown an impressive performance, they still lack in several important aspects such as robustness, systematic generalization, interpretability, reasoning, and the ability to create new knowledge from limited experience. To address these limitations, learning representations of data that capture its underlying hidden structure is thought to be crucial. This dissertation develops architectures and algorithms for learning structure-aware representations without any human supervision or labels, aiming to capture (1) the 4D spatiotemporal structure and (2) the compositional object-centric structure of visual scenes. In Part One, we learn the 4D spatiotemporal structure by integrating observations over time and employing predictive coding. This improves novel view synthesis over baselines, pointing to a superior representation of the underlying scene geometry and dynamics. In Part Two, we introduce a novel object-centric representation learning method by inverting a flexible decoder, demonstrating for the first time the ability to decompose complex scene images and render systematically novel images. We further extend this method to handle videos through two routes: a sequential route via recurrence and a parallelizable route via causal attention. In Part Three, we shift focus towards systematic generalization and concept reuse. We develop a novel method that not only disentangles objects but also learns intra-object factor concepts that are optimized for reusability across scenes. Lastly, we develop a novel image-to-image translation benchmark to measure the ability of deep learning models to generalize to systematically novel visual scenes. Location: CoRE 305 Committee: Professor Sungjin Ahn (Chair) Professor Hao Wang Professor Yongfeng Zhang |
| Start Date: 03 Dec 2024; Start Time: 12:30PM - 02:00PM Title: Production, Consumption, Absorption, Impact of News Bio: David Rothschild is a Senior Principal Research and Economist at Microsoft Research. His work pushes the boundaries on varying data and methods: polling, prediction markets, social media and online data, large behavioral and administrative data, and large-language models. His work focuses on solving practical and interesting questions including: mapping and updating public opinion, the market for news, effect of advertising, behavioral economics and finance, an economist's take on public policy, and how AI is affecting all of this. Speaker: Abstract: In this talk, David will discuss an ongoing project using LLM and Human-in-the-Loop coding to label news in near-real time (https://mediabiasdetector.seas.upenn.edu/) And discuss key findings about what mainstream media chose to cover in 2024 and how that may have impacted the general voting population. Location: CoRE 301 Committee: |
| Start Date: 06 Dec 2024; Start Time: 02:00PM - 04:00PM Title: Conceptual Explanations for Vision and Language Foundation Models Bio: Speaker: Abstract: Vision and language foundation models, such as Vision Transformers (ViTs) and Pretrained Language Models (PLMs), have seen significant advances due to their capability to process and understand visual and textual information. However, trustworthy and interpretable explanation methods for these models remain underdeveloped, especially in post-hoc conceptual explanations that span multiple modalities. Our work introduces a unified framework to generate conceptual explanations for vision and language models, addressing core desiderata, such as faithfulness and stability. Specifically, we introduce variational Bayesian conceptual explanation methods that model the latent distributions of visual/textual token embeddings, providing post-hoc explanations at the dataset, image, and patch levels. Our analysis reveals how modeling multi-level joint distributions of visual or language embeddings can offer interpretable insights, bridging the gap between vision-language model predictions and human-understandable concepts. Extensive experiments across various benchmarks demonstrate that our approach consistently outperforms existing explanation methods by meeting these desiderata and offering a comprehensive analysis of model predictions. Location: CoRE 305 Committee: Professor Hao Wang (Chair) Professor Dimitris Metaxas Professor Yongfeng Zhang Professor Sharon Levy |
| Start Date: 12 Dec 2024; Start Time: 10:00AM - 12:00PM Title: Bridging the Gap Between High-Level Quantum Algorithms and Low-Level Quantum Assembly: Qubit Reuse and Compilation Optimization Bio: Speaker: Abstract: Despite the rapid advancements in quantum computing, practical implementation is hindered by challenges such as limited qubit resources, low fidelity, and error-prone operations. This dissertation presents a comprehensive exploration of quantum compilation and error mitigation through three key projects. First, the CaQR framework leverages mid-circuit measurements and qubit reuse to address qubit limitations, reducing resource constraints, and enhancing fidelity on real quantum hardware. Second, AutoBraid introduces compiler-level support for surface code error correction, enabling more efficient fault-tolerant quantum computation. Lastly, QASMTrans, an open-source quantum compiler, supports scalable algorithms like QAOA and achieves significant performance improvements across diverse quantum architectures. Together, these contributions bridge the gap between high-level quantum algorithms and low-level hardware, advancing scalable, efficient, and error-resilient quantum computing. This work lays the foundation for future research, pushing the boundaries of practical quantum computation. Location: CoRE 305 Committee: Associate Professor Eddy Z. Zhang Distinguished Professor Mario Szegedy Assistant Professor Yipeng Huang |
| Start Date: 12 Dec 2024; Start Time: 01:00PM - 02:30PM Title: Tabletop Object Rearrangement: Structure, Complexity, and Efficient Combinatorial Search-Based Solutions Bio: Speaker: Abstract: This thesis provides an in-depth structural analysis and efficient algorithmic solutions for tabletop object rearrangement with overhand grasps (TORO), a foundational task in advancing intelligent robotic manipulation. Rearranging multiple objects in a confined workspace presents two primary challenges: sequencing actions to minimize pick-and-place operations—an NP-hard problem in TORO—and determining temporary object placements (“buffer poses”) within a cluttered environment, which is essential yet highly complex. For TORO with available external free space, this work investigates the minimum buffer space, or “running buffer size,” required for temporary relocations, presenting both theoretical insights and exact algorithms. For TORO without external free space, the concept of lazy buffer verification is introduced, with its efficiency evaluated across various manipulator configurations, including single-arm, dual-arm, and mobile manipulators. Location: SPR-402, 1 Spring St, New Brunswick Committee: Professor Jingjin Yu Professor Kostas Bekris Associate Professor Abdeslam Boularias Associate Professor Kaiyu Hang (external) |
| Start Date: 13 Dec 2024; Start Time: 01:00PM - 02:30PM Title: Learning Differentiable Tensegrity Dynamics with Graph Neural Networks Bio: Speaker: Abstract: Tensegrity robots are composed of rigid struts and flexible cables. They constitute an emerging class of hybrid rigid-soft robotic systems and are promising systems for a wide array of applications, ranging from locomotion to assembly. They are difficult to control and model accurately, however, due to their compliance and high number of degrees of freedom. To address this issue, prior work has introduced a differentiable physics engine designed for tensegrity robots based on first principles. In contrast, this work proposes the use of graph neural networks to model contact dynamics over a graph representation of tensegrity robots, which leverages their natural graph-like cable connectivity between end caps of rigid rods. This learned simulator can accurately model 3-bar and 6-bar tensegrity robot dynamics in simulation-to-simulation experiments where MuJoCo is used as the ground truth. It can also achieve higher accuracy than the previous differentiable engine for a real 3-bar tensegrity robot, for which the robot state is only partially observable. When compared against direct applications of recent mesh-based graph neural network simulators, the proposed approach is computationally more efficient, both for training and inference, while achieving higher accuracy. Location: Room 402, 4th floor, 1 Spring Street, Downtown New Brunswick Committee: Assistant Professor Mridul Aanjaneya Professor Kostas Bekris Associate Professor Abdeslam Boularias Assistant Professor He Zhu |
| Start Date: 16 Dec 2024; Start Time: 10:00AM - 11:30AM Title: Computational Modeling for Food Image and Video Understanding Bio: Speaker: Abstract: Although sophisticated computer vision models have excelled in numerous tasks, they often struggle with images or videos related to food, especially those depicting cooked meals or food preparation. The food domain presents numerous challenges due to factors such as hidden ingredients, alterations in the appearance of ingredients during cooking, and significant variability in images of dishes prepared with the same recipes. Moreover, cooking procedures frequently allow for interchangeable or potentially parallel sequences of steps, leading to significant variability in the videos depicting cooking. Therefore, designing computer vision models for applications involving food images and videos necessitates particular attention because of the domain’s inherent complexity.This dissertation focuses on computational modeling for both cooking images and videos. For images, we study the problem of predicting the relative amounts of the ingredients. We propose PITA framework leveraging retrieval features and Wasserstein distance to solve the problem. We show PITA significantly outperforms previous baselines. In the context of videos, we address the challenge of deriving an instruction flow graph for each video, which models sequencing as well as parallelism or interchangeability of the cooking steps. We propose Box2flow, which combines features from object detection and BERT to calculate pairwise edge probabilities. A spanning-tree algorithm is then used to predict the flow graph from edge probabilities for each video input. We show that Box2flow outperforms standard captioning-based methods.Our research paves the way for numerous important directions in future studies, particularly when integrated with Large Language Models. The methodologies developed here for food analysis are adaptable and could be utilized in broader computer science challenges involving complex datasets. Location: CBIM 22 Committee: Prof. Vladimir Pavlovic (chair) Prof. Yongfeng Zhang Prof. Hao Wang |
| Start Date: 17 Dec 2024; Start Time: 10:00AM - 12:00PM Title: Enhancing Quantum Computing Efficiency: Compilation Strategies Leveraging Algorithmic and Hardware Insights Bio: Speaker: Abstract: Quantum computing has rapidly advanced, with diverse quantum devices such as superconducting qubits, trapped ions, neutral atoms, and photonic chips. Since Richard Feynman's 1981 proposal, significant algorithms—including Shor's algorithm, Grover's search, and Variational Quantum Algorithms (VQAs)—have been developed, underscoring the need for efficient systems that bridge high-level algorithms and hardware implementations. Quantum algorithms, typically expressed in high-level languages, are transformed into logical circuits, then mapped onto physical circuits using hardware-specific basis gates via qubit mapping and routing, and finally executed through control pulses. Future quantum systems are expected to incorporate error correction codes to enhance computational reliability.My research focuses on algorithm-specific compilation with cross-stack optimization to enhance the efficiency of quantum program execution on existing hardware. I explore optimization opportunities arising from gate commutativity in algorithms like the Quantum Approximate Optimization Algorithm (QAOA) and Quantum Fourier Transform (QFT), as well as flexibility in circuit synthesis for Variational Quantum Eigensolver (VQE) algorithms. Additionally, I analyze often-overlooked hardware characteristics, such as the regularity of qubit connectivity in modern quantum devices, to inform compilation strategies.By studying gate commutativity and qubit connectivity, we discovered a compilation pattern for QAOA achieving linear circuit depth for clique graphs. Building upon this, we developed a general framework adaptable to practical cases, effectively handling sparsity of problem graphs and hardware noise variability. This led to up to 72% reduction in circuit depth and 66% reduction in gate count on IBM and Google architectures with up to 1,024 qubits, outperforming baselines in experiments on IBM Mumbai.We extended this to QFT compilation, resulting in the first linear-depth QFT circuits for architectures like Google Sycamore, IBM heavy-hex, and 2D grids with arbitrary qubit counts. Our methods overcome limitations of techniques relying on SAT solvers or heuristics, which often suffer from long compilation times or suboptimal outcomes due to large search spaces.In another contribution, we introduced Tetris, a compilation framework for VQA applications. Tetris focuses on reducing two-qubit gates during compilation, as these have higher error rates and execution times. By exploiting opportunities in circuit synthesis and using a refined intermediate representation of Pauli strings, Tetris reduces two-qubit gate counts and mitigates hardware mapping costs through a fast bridging approach. Overall, Tetris achieves reductions of up to 41.3% in CNOT gate counts, 37.9% in circuit depth, and 42.6% in circuit duration across molecular simulations compared to state-of-the-art approaches.The methodologies and insights from my research are not limited to these three scenarios; they can be applied to future quantum program compilation tasks. By focusing on cross-stack optimization and leveraging both algorithmic properties and hardware characteristics, my work contributes to bridging the gap between quantum algorithms and hardware, significantly improving the efficiency and scalability of quantum computing implementations. Location: CoRE 305 Committee: Professor Zheng Zhang Professor Mario Szegedy Professor Yipeng Huang |
| Start Date: 30 Jan 2025; Start Time: 01:30PM - 03:00PM Title: Efficient Near-Duplicate Text Alignment Search via Bottom-k Sketches Bio: Speaker: Abstract: The near-duplicate text alignment search problem—aimed at identifying all near-duplicate passage pairs between a query document and a collection of source documents—presents a computationally intensive challenge. This task is crucial for applications like plagiarism detection and LLM memorization. Brute-force methods require comparing O(n^2 m^2) passage pairs between a single query document with n words and a single source document with m words, making them impractical for large corpora. Existing solutions often rely on heuristics like "seeding-extension-filter," which lack theoretical guarantees and involve numerous difficult-to-tune hyperparameters. A previous work ALLIGN employs min-hash sketches to solve this problem. However, it is limited to comparisons between pairs of documents.To address these limitations, we propose leveraging bottom-k sketching to estimate Jaccard similarity between passages. A key insight in our approach is that contiguous passages share the same bottom-k sketch. We introduce the concept of “compact window” to represent all passages that share the same sketch in a concise manner and develop an efficient algorithm to group these passages in O(nlogn+nk)time, reducing the total number of sketches from O(n^2) to O(nk) for a document with n words. Experimental results on real-world datasets demonstrate that our techniques achieve high efficiency.List of Publication(s):TxtAlign: Efficient Near-Duplicate Text Alignment Search via Bottom-k Sketches for Plagiarism Detection (SIGMOD2022) Location: CoRE 305 Committee: Assistant Professor Dong Deng (Advisor) Associate Professor Yongfeng Zhang Assistant Professor He Zhu Assistant Professor Arpita Biswas |
| Start Date: 04 Feb 2025; Start Time: 10:30AM - 12:00PM Title: On The Myths of Reasoning Language Models: What is Next? Bio: Dan Roth is the Eduardo D. Glandt Distinguished Professor at the Department of Computer and Information Science, University of Pennsylvania and the Chief AI Scientist at Oracle. Until June 2024 Dan was a VP/Distinguished Scientist at AWS AI. In his role at AWS Roth led over the last three years the scientific effort behind the first-generation Generative AI products from AWS, including Titan Models, Amazon Q efforts, and Bedrock, from inception until they became generally available. Dan is a Fellow of the AAAS, ACM, AAAI, and ACL. In 2017, Dan was awarded the John McCarthy Award; he was recognized for “for major conceptual and theoretical advances in the modeling of natural language understanding, machine learning, and reasoning.” He has published broadly in natural language processing, machine learning, knowledge representation and reasoning, and learning theory, was the Editor-in-Chief of the Journal of Artificial Intelligence Research (JAIR) and has served as a Program Chair and Conference Chair for the major conferences in his research areas. Roth has been involved in several startups; most recently he was a co-founder and chief scientist of NexLP, a startup that leverages the latest advances in Natural Language Processing, Cognitive Analytics, and Machine Learning in the legal and compliance domains. NexLP was acquired by Reveal. Dan received his B.A Summa cum laude in Mathematics from the Technion, Israel and his Ph.D. in Computer Science from Harvard University in 1995. Speaker: Abstract: The rapid progress made over the last few years in generating linguistically coherent natural language has blurred, in the mind of many, the difference between natural language generation, understanding, and the ability to reason with respect to the world. Nevertheless, robust support of high-level decisions that depend on natural language understanding, and one that requires dealing with “truthfulness” are still beyond our capabilities, partly since most of these tasks are computationally more complex than language models can support, are very sparse, often require grounding, and may depend on new types of supervision signals. I will discuss some of the challenges underlying reasoning and argue that we should focus on LLMs as orchestrators – coordinating and managing multiple models and special purpose agents. I will discuss some of the challenges and present some of our work in this space, focusing on supporting planning and a range of quantitative, visual, and spatial reasoning tasks. Location: CoRE 301 Committee: |
| Start Date: 07 Feb 2025; Start Time: 04:00PM - 06:00PM Title: Controlling Long-Horizon Behavior in Reinforcement Learning Bio: Speaker: Abstract: Reinforcement learning (RL) has been successful in a very wide range of domains, from Atari games, to robotic control, to theorem proving. However, existing reinforcement learning methods tend to only work well over short time horizons -- many aspects of the problem become exponentially harder as time horizons grow. In particular, long time horizons make exploration and planning very challenging, and can worsen biases in some popular algorithms. In hard-exploration settings, standard algorithms like MCTS can take exponential or super-exponential time to explore the full environment. In model-based RL, predictions with a learned model may become exponentially worse as the time horizon increases. In the Hindsight Experience Replay algorithm (HER), biases from stochastic environments become worse over longer time horizons. The common problem underlying these issues is that local information available to the policy is not always sufficient to solve the problem -- global information about the state space must sometimes be made available to the agent to enable a efficient solution. This thesis aims to address these issues by developing mathematical tools for representing and controlling the long-horizon behavior of RL agents. In particular, focus is placed on the state occupancy measure and state successor measure. These tools are useful because they reduce RL to maximizing a linear objective with linear constraints. This means that many problems can be reformulated in a way that requires far less dependence on the time horizon. In a wide range of domains, I show that this tool can be used to address issues with convergence, exploration, and bias. Location: Room 402, 4th floor, 1 Spring Street, Downtown New Brunswick Committee: Associate Professor Abdeslam Boularias Assistant Professor Hao Wang Professor Kostas Bekris External member: Bo Yuan |
| Start Date: 11 Feb 2025; Start Time: 10:30AM - 12:00PM Title: Our Journey Towards a Diverse Computing Program: What Worked, Where we Are, and What we have Learned Bio: Dr. Ana Paula Centeno is an Associate Teaching Professor in the Department of Computer Science at Rutgers, New Brunswick. She received her PhD in 2019 from Rutgers and as a teaching faculty at Rutgers, has been involved in research to understand the pathway blocks that impede student success in the computer science major. She has published papers on this topic, has been active with the Advancing Women in Computer Science (AWiCS) initiative, and has received the Rutgers Presidential Employee Excellence Recognition Award and the Provost’s Award for Excellence in STEM Diversity in 2023 and 2021, respectively. Speaker: Abstract: Women and individuals from underrepresented groups are significantly underrepresented in educational pathways and professional roles in computing. This imbalance not only reveals broader social inequities but also alters the viewpoints of both current and future technologists. These perspectives can lead to a lack of motivation or self-efficacy in academia and implicit hiring and algorithmic biases in the professional world.The underrepresentation of women and other marginalized groups presents a significant challenge. Addressing the barriers to entry and advancement in the field is essential for creating a more inclusive computing landscape that reflects and serves the needs of all people. I will discuss our work toward addressing these barriers at the computer science department at Rutgers - report the approaches that we implemented, starting Fall 2019, their outcomes, and the challenges faced throughout this journey's stage. More importantly, we show that the changes implemented in our program had a positive impact on both underrepresented groups and the overall student body. Location: CoRE 301 Committee: |
| Start Date: 07 Mar 2025; Start Time: 01:00PM - 02:00PM Title: Bayesian Optimization for Drug Discovery: Tackling High-Dimensions and Constrained Multi-Objectives Bio: Chong Liu is an Assistant Professor of Computer Science at University at Albany, State University of New York. His research interests lie broadly in machine learning, optimization, and AI for science, with emphasis on Bayesian optimization, bandit algorithms, active learning, and drug discovery. He is an area chair of ICML and AISTATS, associate editor of IEEE-TNNLS, and editorial board reviewer of JMLR. He organized the NeurIPS-2023 Workshop on New Frontiers of AI for Drug Discovery and Development (AI4D3-2023). He received Ph.D. in Computer Science from UC Santa Barbara in 2023 and spent one year as a Data Science Institute Postdoctoral Scholar at the University of Chicago. Speaker: Abstract: Drug discovery is an expensive process that involves sequentially screening and examining a large pool of candidate molecules. Due to its black-box nature, Bayesian Optimization (BO) is usually applied in this task. In practice, candidate molecules may be represented in high-dimensional spaces and a successful candidate is expected to hit multiple target performance criteria. Unfortunately, existing BO work usually falls short in these scenarios, which prevents more practical applications in drug discovery. In the first part of this talk, I’ll show how to solve BO without classical Gaussian process assumption but instead with parametric function approximation. With this new technique, we can prove a new input dimension-free cumulative regret bound. In the second part, I’ll show how to solve the constrained multi-objective BO through optimistic constraints estimation. Both theoretical and empirical results will be reported, and future directions will be discussed. Location: CoRE 301 Committee: |
| Start Date: 10 Mar 2025; Start Time: 10:30AM - 12:00PM Title: Regularization in Deep Neural Networks Bio: Dr. Bernhard Firner's career has focused on two main interests: deep learning and low-power embedded systems. His recent work has focused on the intersection of those topics: real-time control of embedded systems with onboard neural networks. After completing his PhD at Rutgers in 2014, Bernhard joined a newly created group in NVIDIA to work on end-to-end learning for autonomous vehicles. His work resulted in numerous patents, popular academic papers, and notable technology demonstrations. Within five years, his team's autonomous vehicles could achieve average distances of 500km between failures. These results were on public U.S. highways in all lighting and weather conditions, including snow. Bernhard's experience includes time with startup companies. In one case developing low-power wireless sensors that ran for 20 years on coin cell batteries and in the other creating neural networks for real-time control of autonomous drones. Earlier in his career, he developed software for real-time embedded avionics platforms at a well-established avionics company. Speaker: Abstract: You may have heard that overfitting is a problem in machine learning. You may have even heard that regularization fixes the problem. But what is regularization?Regularization techniques pre-date modern machine learning, including deep neural networks. Although deep neural networks are surprisingly robust to overfitting, regularization is still an essential part of neural network training. In this talk, we will look at overfitting and regularization in neural networks. How do they resist overfitting? What techniques can we use to regularize neural networks? And what are some of the problems that regularization solves? Location: CoRE 301 Committee: |
| Start Date: 11 Mar 2025; Start Time: 10:00AM - 12:00PM Title: Learning Structured Representations from Images and Videos Bio: Speaker: Abstract: Our visual world is highly compositional and modular, consisting of objects, parts, and attributes that are organized in a hierarchical and relational manner. This structure allows for complex scenes to be explained by combinations of simpler elements, leading to better interpretability and facilitating systematic generalization. While humans innately view the world in terms of objects, training a neural network to learn these concepts from raw visual input remains a challenge. This dissertation presents several advancements in learning structured, object-centric representations from images and videos. The first part of the dissertation investigates model architectures for learning unsupervised object-centric representations. We introduce SPACE, a method that uses parallel-spatial attention to efficiently decompose multi-object images into factored object representations, separating the location (bounding box) of objects from their appearance representations. We then investigate using inverted-attention transformers for object-centric learning, identifying the minimal set of changes to a standard Transformer architecture that can enable object discovery in images and videos. The second part of the dissertation explores how object-centric representations can be used as a natural way to tokenize visual input for the modern transformer architecture. We introduce the Object-Centric Video Transformer (OCVT) that uses SPACE representations as tokens in a generative video transformer, enabling the generation of videos with long-term dependencies and learning representations that are useful for downstream reasoning tasks. Lastly, we present Neural Language of Thought Models (NLoTM), which learns hierarchical and composable discrete representations that are aligned with objects and their properties. These discrete representations can then be used to train an autoregressive transformer-based prior to obtain an object-centric density model. Sampling from this prior allows us to generate an image one object at a time, based on its properties. Location: Virtual on Zoom Committee: Professor Sungjin Ahn (advisor) Professor Abdeslam Boularias Professor Vladimir Pavlovic |
| Start Date: 11 Mar 2025; Start Time: 10:30AM - 12:00PM Title: Machine Learning for Formal Software Verification Bio: Emily First is a postdoctoral researcher at UC San Diego working with Professor Sorin Lerner. She previously completed her PhD at UMass Amherst, where she was awarded the UMass CICS Outstanding Dissertation Award. Her research is at the intersection of software engineering, programming languages, and machine learning. She focuses on creating tools to automatically generate proofs of software correctness. Her work has also received two distinguished paper awards. Speaker: Abstract: Formal verification using proof assistants is an effective way of improving software quality, but is expensive. Recent work has seen a wave of neural theorem provers, which use neural models to predict proofs and perform proof search to traverse the space of possible proofs. This talk presents my work on building proof search approaches for LLMs, focusing on what inputs LLMs need and how to better guide their search. I'll conclude my talk with a vision of how LLM-based proof search and, more broadly, the formal software verification pipeline can benefit from human-AI collaboration. Location: CoRE 301 Committee: |
| Start Date: 11 Mar 2025; Start Time: 12:00PM - 01:00PM Title: Advancing Knowledge-Intensive Tasks in NLP: Entity Linking and Beyond Bio: Speaker: Abstract: Knowledge-intensive NLP tasks—such as open-domain question answering, fact-checking, and entity linking—rely on large external knowledge sources to handle the complexity of real-world problems. Despite significant progress, existing methods remain suboptimal and leave ample room for improvement. In this talk, I focus on entity linking as a prime example of these challenges. Entity linking involves identifying mention spans within a document and associating each span with the correct entity in a knowledge base. I will introduce our novel approach to entity linking, which not only achieves strong results but also overcomes key shortcomings of previous models. In addition, this method enables synergy with open-domain question answering, highlighting how progress in one knowledge-intensive task can benefit others. Building on this foundation, I will then briefly discuss how we can enhance the core retrieval component underpinning all knowledge-intensive tasks through a unified, multitask-trained model. Finally, I will outline promising directions for future work, with the broader goal of advancing knowledge-intensive NLP in a more general and integrative manner.A list of publications exam will be based on: Understanding hard negatives in noise contrastive estimation. NAACL 2021.EntQA: Entity linking as question answering. ICLR 2022.Seq2seq is all you need for coreference resolution. EMNLP 2023.Improving Multitask Retrieval by Promoting Task Specialization. TACL 2023.Google scholar: https://scholar.google.com/citations?hl=zh-CN&user=yWEhrEAAAAAJ Location: CoRE 305 Committee: Assistant Professor Karl Stratos Assistant Professor Hao Wang Associate Professor Abdeslam Boularias Professor of Professional Practice James Abello Monedero |
| Start Date: 13 Mar 2025; Start Time: 10:30AM - 12:00PM Title: Building Shapeable and Reliable AI with Probabilistic Modeling Bio: Jake Snell is a postdoctoral researcher at Princeton University working with Tom Griffiths. He earned his bachelor's degree in Biomedical Engineering at Yale University and his Ph.D. in Computer Science at the University of Toronto, advised by Richard Zemel. His research focuses on integrating deep learning and probabilistic modeling to build AI systems that are more reliable and easier to control. He is a recipient of the Schmidt DataX Postdoctoral Fellowship and finalist for best student paper award at the IEEE International Conference on Image Processing in 2017. Speaker: Abstract: Current LLM architectures are exceedingly powerful yet remain difficult to control while unexpectedly producing harmful outputs such as hallucinations and toxic content. In this talk, I will show how deep learning can borrow the strengths of probabilistic models to address these issues. First, I will show how probabilistic models influence the way that deep neural networks generalize via metalearning. Second, I will demonstrate how probabilistic inference facilitates safe deployment of AI systems in risk-sensitive settings by producing rigorous guarantees about their performance. I will conclude with future directions integrating these lines of work to build deep learning algorithms with capabilities that are verifiable by design. Location: CoRE 301 Committee: |
| Start Date: 14 Mar 2025; Start Time: 10:30AM - 12:00PM Title: Leveraging the Wisdom of Clouds for Internet Security Bio: Eric Pauley is a Ph.D. candidate at the University of Wisconsin–Madison, advised by Patrick McDaniel. His research interests encompass data-driven approaches to evaluating and improving the security of networked software systems, with a particular focus on cloud computing. His work has led to practical improvements in the security of cloud-based systems through both remediations by major providers and services offered by his company, DScope Security. His research in security measurement has earned best paper runner-up at the ACM Internet Measurement Conference, a finalist spot in the CSAW Applied Research Competition, and the UW–Madison Computer Sciences Outstanding Graduate Researcher Award. Eric is also an avid backpacker, instrument-rated private pilot, and birder. Speaker: Abstract: Over the past decade, networked systems have consolidated under just a handful of hyperscale cloud providers (e.g., AWS, Azure). While this offers logistical and economic advantages, attackers specifically target providers and their customers, a shift that has left traditional network vantage points blind to the most sophisticated adversaries. In this talk, I’ll explore how we adapt Internet measurement to these new deployment models to regain situational awareness and defend modern service deployments. I’ll introduce DScope, a new Internet telescope that continuously relocates its vantage point across public cloud infrastructure. Unlike prior approaches that use a fixed vantage point, this allows us to observe the most sophisticated attackers that actively avoid existing measurement infrastructure. Our dynamic approach also achieves a statistically representative view of cloud-based attacks, a property that we prove for the first time.Using data from DScope, I’ll also discuss how the shared networking environment of public clouds leads to new vulnerabilities. We’ll examine the problem of latent configuration, which occurs when cloud customers reference network resources that are then reused by other tenants. This new security risk is uniquely enabled by public clouds, but through rigorous analysis and systems design we can make cloud deployments more secure in practice. I’ll conclude by discussing open problems and future work in leveraging Internet vantage points for security, with a focus on intelligent interactivity and rapid response to emergent threats. Location: CoRE 301 Committee: |
| Start Date: 17 Mar 2025; Start Time: 10:30AM - 12:00PM Title: A holistic view of Internet security Bio: Dr. Wang is an Associate Research Scholar at Princeton University. He received his Ph.D. in Computer Science from the University of Wisconsin-Madison in 2018. His research interests lie in network security and privacy. His research focuses on using data-driven approaches to understand how major Internet components function in practice, identifying cross-boundary attacks, and developing practical and incrementally deployable defenses. His work was awarded the Caspar Bowden Award for Outstanding Research and Best Student Paper Runners-up at PETS 2022, and has been recognized in various ways (adoptions, invited talks, etc.) by industry leaders such as Google, Cloudflare, and Intel. Speaker: Abstract: The Internet ecosystem consists of various interconnected infrastructures, protocols, and services that depend on each other for seamless operation. Despite rigorous security analyses of individual components, security challenges can arise from interactions between components, necessitating a holistic approach to Internet security. This talk will explore security threats affecting major Internet components and the innovative defenses developed to mitigate these challenges. Specifically, I will demonstrate how cross-boundary attacks, such as traffic analysis, exploit encrypted network traffic to compromise the security and privacy of applications at higher layers, and introduce a novel defense mechanism to counter these risks. Additionally, I will discuss cross-layer routing attacks on Public Key Infrastructure (PKI), highlighting a practical defense that was successfully deployed at Let’s Encrypt and Google since 2020 and became an Internet standard in 2024. Through these case studies, I will illustrate the evolving landscape of Internet security and the importance of holistic security analysis. Location: CoRE 301 Committee: |
| Start Date: 21 Mar 2025; Start Time: 11:00AM - 01:00PM Title: Conditional Generative Modeling for Holistic Human Behaviors Bio: Speaker: Abstract: Animating behaviors for virtual humans and digital characters has immense potential for creating immersive and entertaining metaverse experiences in computer graphics and virtual reality. However, it remains challenging to craft lifelike human behaviors because natural human performance is inherently a synchronization of multiple modalities, and while existing approaches can generate individual aspects of human behavior, they often fail to capture the nuanced interplay between emotions, movements, and social interactions. Our approach to addressing these challenges is conditional generative modeling, which aims to synthesize human behaviors with high fidelity, achieve synchronization, and provide controllability for the performance.The first part of this dissertation advances the understanding of conversational behavior generation through novel emotional modeling. We begin by identifying critical limitations in existing embodied conversational agents, particularly their inability to maintain affect consistency across different behavioral modalities. Based on these findings, we propose an emotion-conditioned generative model that successfully disentangles content and emotion from input speech, enabling the generation of emotionally coherent facial animations. Our framework demonstrates significant improvements in both the naturalness and emotional expressiveness of virtual human performances compared to previous methods. The second part extends our conditioning approach to broader human movements and social interactions. We introduce CASIM, a semantic injection mechanism that fundamentally reimagines text-to-motion generation by incorporating the composite nature of human motions in both spatial and temporal domains. Our extensive experiments on the HumanML3D and KIT benchmarks demonstrate substantial improvements in text-motion correspondence and motion quality across multiple state-of-the-art models. Furthermore, we pioneer the generation of group activities by developing a novel diffusion-based framework with an interaction transformer that models inter-person dynamics, capable of synthesizing socially interacting groups of arbitrary sizes. This includes creating the first comprehensive dataset and evaluation metrics for group activity generation, establishing new benchmarks for this emerging research direction. Location: CoRE 301 Committee: Professor Mubbasir Kapadia Professor Vladimir Pavlovic Professor Dimitris Metaxas |
| Start Date: 24 Mar 2025; Start Time: 10:30AM - 12:00PM Title: Formal Methods for Trustworthy AI and Autonomy Bio: Pavithra Prabhakar is professor in the department of computer science, and the Peggy and Gary Edwards Chair in Engineering at Kansas State University. She obtained her doctorate in computer science and a masters degree in applied mathematics from the University of Illinois at Urbana-Champaign, followed by a CMI postdoctoral fellowship at the California Institute of Technology. Prior to coming to K-State, she spent four years at the IMDEA Software Institute in Spain as a tenure-track assistant professor. Prabhakar’s expertise is on AI-enabled autonomous, cyber-physical, and robotic systems, with applications that span aerospace, automotive, and agricultural automation. She has authored over 100 peer-reviewed scientific articles, with several receiving invited presentations and best paper awards, leading KSU Computer Science Department to within top 20 in the area of embedded and real-time systems as reported by csrankings.org. She is the recipient of a Marie Curie Career Integration Grant from the European Union (2014), an NSF CAREER Award (2016), an ONR Young Investigator Award (2017), NITW distinguished young alumnus award (2021), an Amazon Research Award (2022), and a CRA Future Leader recognition (2024). Currently, Dr. Prabhakar is serving the National Science Foundation as a Program Director in the Computer and Information Science and Engineering (CISE) Directorate as an IPA. In this role, she oversees and manages programs supporting research across Cyber-Physical Systems, Robotics, AI, and Formal Methods, handling a portfolio of over 200 projects and principal investigators with a total budget exceeding $100 million. She leads the Formal Methods in the Field (FMitF) program, is a founding program director for the Safe Learning Enabled Systems (SLES) program and is a cognizant program director for the Cyber-Physical Systems (CPS) and Foundational Research in Robotics (FRR) programs which are cross-cutting initiatives spanning the CISE and Engineering directorates. In addition, she oversees the formal methods and verification portfolio of the core program in the Software-Hardware Foundations cluster. Speaker: Abstract: AI-based components have become an integral part of Cyber-Physical Systems (CPS) enabling transformative functionalities. With the ubiquitous use of machine learning components in perception, control and decision making in safety critical application domains such as automotive and aerospace, rigorous analysis of these systems has become imperative toward real-world deployment. In this talk, we will present an overview of the work in our group on formal approaches to designing and certifying AI-enabled CPS supported by NSF, Amazon, USDA and ONR research grants.First, we consider formal verification techniques for AI-enabled CPS modeled as a neural network controlled-dynamical system. One of the main challenges is the scalability of theverification algorithms to complex dynamics and large neural networks. One research direction consists of abstraction algorithms that reduce the size of the system to scale up verification. We will discuss our work on interval neural networks and interval image data structures for abstraction-based analysis of safe AI-enabled CPS. Our experimental results demonstrate the benefits of our methods in analyzing large scale AI-enabled CPS. We will discuss other foundational questions including providing certificates for evolving neural network components, and properties beyond safety.Next, we present our work on the applications of formal verification and synthesis to robotics, aerospace systems and agricultural automation. Specifically, we will discuss a switched system framework aimed at the design of intelligent and safe aerospace systems. In addition, we will discuss the design of multi-agent path planning and coverage planning algorithms with applications in agricultural automation. Finally, we will present the vision for a research agenda in trustworthy AI-enabled CPS. Location: CoRE 301 Committee: |
| Start Date: 26 Mar 2025; Start Time: 02:00PM - 03:00PM Title: Autoregressive Action Sequence Learning for Robotic Manipulation Bio: Speaker: Abstract: A key challenge remains to design a universal policy architecture that performs well across diverse robots and task configurations. In this work, we address this by representing robot actions as sequential data and generating actions through autoregressive sequence modeling. Existing autoregressive architectures generate end-effector waypoints sequentially as word tokens in language modeling, which are limited to low-frequency control tasks. Unlike language, robot actions are heterogeneous and often include high-frequency continuous values---such as joint positions, 2D pixel coordinates, and end-effector poses—which are not easily suited for language-based modeling. Based on this insight, we extend causal transformers' single-token prediction to support predicting a variable number of tokens in a single step through our Chunking Causal Transformer (CCT). This enhancement enables robust performance across diverse tasks of various control frequencies, greater efficiency by having fewer autoregression steps, and lead to a hybrid action sequence design by mixing different types of actions and using a different chunk size for each action type. Based on CCT, we propose the Autoregressive Policy (ARP) architecture, which solves manipulation tasks by generating hybrid action sequences. We evaluate ARP across diverse robotic manipulation environments, including Push-T, ALOHA, and RLBench, and show that ARP, as a universal architecture, matches or outperforms the environment-specific state-of-the-art in all tested benchmarks, while being more efficient in computation and parameter sizes.List of Publications:Autoregressive action sequence learning for robotic manipulation. (RA-L 2025). Xinyu Zhang, Yuhan Liu, Haonan Chang, Liam Schramm, Abdeslam Boularias Location: Room 402, 4th floor, 1 Spring Street, Downtown New Brunswick Committee: Professor Abdeslam Boularias Professor Kostas Bekris Professor He Zhu Professor Karthik C. S. |
| Start Date: 27 Mar 2025; Start Time: 10:30AM - 12:00PM Title: Pushing the Boundaries of Modern Application-Aware Computing Stacks Bio: Christina Giannoula is Postdoctoral Researcher at the University of Toronto, working with Professors Gennady Pekhimenko, Andreas Moshovos and Nandita Vijaykumar. She is an affiliated senior researcher with the SAFARI research group at ETH Zurich, working with Prof. Onur Mutlu, and is also a senior researcher at CentML company, where she advises research engineers on system design for deep learning. Her research lies at the intersection of computer architecture, systems software and parallel computing, with a focus on high performance, energy efficiency, and programmability for emerging applications. Her current research interests include hardware-software co-design for cutting-edge applications, emerging memory technologies such as processing-in-memory, and systems for machine learning. Christina received her Ph.D. in 2022 from the School of Electrical and Computer Engineering (ECE) at the National Technical University of Athens (NTUA), Greece. She has been recognized with several awards and fellowships for her Ph.D. research, including a three-year Ph.D. Fellowship from the Hellenic Foundation for Research and Innovation (HFRI) and the General Secretariat for Research and Technology (GSRT) and a PhD award (Sep 2021-Oct 2022) from the Foundation for Education and European Culture (IPEP). Moreover, her Ph.D. thesis received the Iakovos Giurunlian NTUA award for the doctoral thesis with the highest industrial impact in 2022. For her Postdoctoral research, she has received research grants from the Vector Institute for Artificial Intelligence and has been recognized as a Rising Star in EECS (Electrical Engineering and Computer Sciences) and a Rising Star in MLSys (Machine Learning Systems) in 2024. Christina also serves as the social media editor for the ACM SIGMICRO organization. For more information, please visit her website at https://cgiannoula.github.io/. Speaker: Abstract: Modern computing systems encounter significant challenges related to data movement data movement in applications, such as data analytics and machine learning. Within a compute node, the physical separation of the processor from main memory necessitates retrieving data through a narrow memory bus. In big-data applications running across multiple nodes, data must be exchanged via narrow network interconnects. This movement of data —both within and across compute nodes— causes significant performance and energy overheads in modern and emerging applications. Moreover, today’s general-purpose computing stacks overlook the particular data needs of individual applications, missing crucial opportunities for untapped performance optimization.In this talk, I will present a cross-stack approach to designing application-aware computing stacks for cutting-edge applications, enabling new synergies between algorithms, systems software, and hardware. Specifically, I will demonstrate how integrating fine-grained application characteristics —such as input features, data access and synchronization patterns— across the layers of general-purpose computing stacks allows for tailoring stack components to meet the application’s specific data needs. This integration enables the stack components to work synergistically to reduce unnecessary or redundant data movement during application execution. I will present a few of my research contributions that propose hardware and software solutions for emerging applications, such as deep learning, and by capitalizing on the emerging processing-in-memory paradigm. Finally, I will conclude by outlining my future plans to design application-adaptive and sustainable computing stacks to significantly enhance performance and energy efficiency in cutting-edge applications. Location: CoRE 301 Committee: |
| Start Date: 28 Mar 2025; Start Time: 10:30AM - 12:00PM Title: Rethinking security for emerging decentralized systems Bio: Mahimna Kelkar is a final-year PhD student at Cornell University. His research broadly focuses on applied cryptography and the foundations of security for blockchains. His work has received recognition through paper awards at CCS’24 and APKC’22 and has been adopted by multiple blockchain companies. Speaker: Abstract: In the past few years, blockchains have emerged as a new class of decentralized systems with wide-ranging applications. But the promise of blockchains has been marred by hype, speculation, and a plethora of high-profile attacks. In this talk, I will demonstrate why blockchain security is challenging---blockchains operate in a radically new, highly adversarial environment where anonymous actors can immediately monetize subtle protocol flaws. I will show powerful new attacks that provide novel insights on two well-studied security primitives---Byzantine consensus within distributed systems, and proofs-of-knowledge within cryptography. Finally, I will discuss why understanding blockchain security has broader utility, and how it can provide a guiding principle when designing secure systems. Location: CoRE 301 Committee: |
| Start Date: 28 Mar 2025; Start Time: 12:00PM - 02:00PM Title: Computer Science Undergraduates Project Showcase Bio: Speaker: Abstract: Location: CoRE 101 Committee: |
| Start Date: 31 Mar 2025; Start Time: 08:30AM - 10:00PM Title: Object-Centric Representation Learning: Methods and Applications Bio: Speaker: Abstract: Deep learning has revolutionized machine learning and computer vision. However, conventional approaches often overlook the inherent compositional and modular nature of the physical world, making it difficult to express modularity, compositionality, and interpretability in the representations. This dissertation addresses this challenge by developing object-centric learning methods that decompose raw visual inputs into individual entities and represent them as modular components. We further demonstrate the effectiveness of these learned representations in applications such as image generation, editing, and visual reasoning. The core contributions of this dissertation include four novel architectures: (1) SCALOR, which employs parallelizable spatial attention to significantly improve scalability in object-centric video understanding; (2) GNM, which integrates distributed and object-centric representations to enable both interpretable representations and density-based generation; (3) LSD, which integrates diffusion models with object-centric learning to handle complex naturalistic scenes and enable real-world image generation and editing applications; and (4) SlotSSMs, which incorporates object-centric principles into state-space models for improved temporal reasoning and long-context video understanding. Together, these contributions advance the field of object-centric learning, addressing critical limitations in existing methods and expanding the applicability of object-centric representations to real-world tasks. Location: CBIM 22 Committee: Professor Sungjin Ahn (advisor/chair) Professor Hao Wang Professor Ruixiang Tang |
| Start Date: 31 Mar 2025; Start Time: 12:30PM - 01:30PM Title: Artificial Intelligence powered Large-Scale Investigation for Advanced Attacks Bio: Speaker: Abstract: Modern cyberattack investigations face significant challenges due to the exponential growth in both the scale and sophistication of attacks. Logs serve as the primary source for attack investigations, but as attacks become increasingly complex and prolonged, traditional log-based investigation struggles with high storage costs, slow query performance, resource-intensive and low precision analysis. This dissertation examines how artificial intelligence (AI) can improve attack investigation systems for large-scale advanced threat analysis by optimizing log storage, management, and analysis within existing investigation pipelines. First, we introduce an AI-driven log compression method that significantly reduces storage demands while preserving full data accessibility. Second, we propose a neural network-based approach for efficiently representing and querying provenance graphs derived from logs, enabling faster and more scalable analysis of system activity causality. Finally, we develop an unsupervised learning framework that automates attack investigation, eliminating the need for manual data labeling and costly graph preprocessing. This framework not only reduces analysis costs but also enhances investigative accuracy by detecting attack patterns across complex, long-term incidents spanning multiple systems. By significantly reducing storage costs and computational overhead while also improving the effectiveness, our research offers security teams better solutions for advanced threat investigation. Location: CoRE 301 Committee: Assistant Professor Dong Deng Shiqing Ma Assistant Professor Sudarsun Kannan |
| Start Date: 01 Apr 2025; Start Time: 10:30AM - 12:00PM Title: Efficient Cryptographic Computation for Real-World Programs and People: Advancing Algorithms and Systems Bio: Yibin Yang is a Ph.D. candidate in Computer Science at Georgia Tech, advised by Vladimir Kolesnikov. Yibin’s research focuses on advancing cryptographic protocols, particularly in Zero-Knowledge Proofs (ZKP) and Secure Multi-Party Computation (MPC). He develops next-generation ZKP and MPC systems that enable seamless execution of real-world off-the-shelf programs over encrypted data, bridging the gap between theoretical cryptography and practical deployment. He is a RSAC Security Scholar and a recipient of Distinguished Paper Awards at ACM CCS. Prior to Georgia Tech, Yibin received his B.Eng. in Computer Science at Tsinghua University in 2019. Speaker: Abstract: Data sharing is indispensable for operational efficiency and groundbreaking innovation in the information age. However, concerns related to intellectual property rights, strict privacy laws, security risks, and complex data custody often hinder this valuable sharing.For nearly four decades, cryptography has — in principle — provided solutions to address most of these concerns. Zero-Knowledge Proofs (ZKP) allow one party, the prover, to prove any property about its private data to another party, the verifier, without disclosing any additional information. More generally, Secure Multi-Party Computation (MPC) enables multiple mutually untrusting parties to jointly compute arbitrary functions on their private inputs, revealing only the intended output. Despite their strong theoretical guarantees, ZKP and MPC deployments remain rare.The core challenges extend beyond a lack of usable tools and systems. They lie in the absence of efficient cryptographic algorithms capable of handling complex, real-world programs, particularly those expressed using high-level programming languages.In this talk, I will explore these challenges and discuss my work that introduces novel cryptographic algorithms to overcome them. I will present end-to-end practically efficient ZKP and 2PC systems built to directly execute off-the-shelf high-level programs within ZKP or 2PC virtual machines. For example, one can prove in ZK interesting properties (e.g. CVE exploitable bugs) of off-the-shelf Linux programs in a few seconds on a laptop.These advances not only enhance the applicability, usability, and adoption of ZKP and MPC but also open new research opportunities in, e.g., compilers, programming languages, and hardware acceleration. Location: CoRE 301 Committee: |
| Start Date: 01 Apr 2025; Start Time: 11:00AM - 12:00PM Title: Near-Duplicate Text Alignment at Scale Bio: Speaker: Abstract: The proliferation of near-duplicate content—text segments differing by minor variations—poses significant challenges in large-scale text corpora, particularly in applications involving large language model (LLM) training and web-scale information retrieval. Studies indicate that 30–45% of web content consists of near-duplicates, leading to inefficiencies such as redundant data storage, increased computational costs, and reduced generalization in LLMs. While exact deduplication techniques, such as suffix arrays, effectively remove identical sequences, the problem of fuzzy subsequence-level deduplication remains largely unsolved due to computational constraints.Existing approaches face two fundamental challenges: (1) combinatorial explosion, where a corpus with n tokens contains O(n^2) subsequences, making exhaustive similarity comparisons computationally prohibitive; and (2) memory-throughput tradeoff, where traditional min-hash-based methods require O(nk) space for k hash functions, rendering them infeasible for trillion-token corpora. Addressing these limitations, we introduce a novel near-duplicate text alignment framework that enables efficient and scalable subsequence-level search.Our contributions are threefold: • One Permutation Hashing (OPH) Compact Windows, a novel compression technique that reduces index size from O(nk) to O(n+k), achieving a 90% reduction in space while maintaining full recall. • Interval-Scan Alignment, an O((n+k)log(n+k)) algorithm leveraging dynamic segment trees to accelerate near-duplicate subsequence detection by 100× over prior methods. • Theoretical Guarantees, providing formal proof that OPH maintains the collision probabilities of traditional min-hash techniques, ensuring no loss in accuracy.Building on this foundation, our ongoing work aims to develop corpus-scale subsequence clustering methods that efficiently group redundant text spans, reducing pairwise comparisons from O(n^2) to O(nk) per document while seamlessly integrating into LLM training pipelines. Our research represents a crucial step toward efficient and scalable fuzzy deduplication, with implications for dataset curation, generative model safety, and large-scale text processing.Publications: • Near-Duplicate Sequence Search at Scale for Large Language Model Memorization Evaluation(SIGMOD 2023) • Near-Duplicate Text Alignment with One Permutation Hashing (SIGMOD 2025) Location: CoRE 305 Committee: Assistant Professor Dong Deng Associate Professor Yongfeng Zhang Associate Professor Desheng Zhang Professor Matthew Stone |
| Start Date: 03 Apr 2025; Start Time: 10:30AM - 12:00PM Title: Learning to Reason with LLMs Bio: Xiang Yue is a Postdoctoral Fellow at Carnegie Mellon University. He received his PhD from The Ohio State University in 2023. His research focuses on understanding and enhancing the reasoning capabilities of large language models. He has been awarded a postdoctoral fellowship from the Carnegie Bosch Institute, two AI rising stars, two Best Paper Finalist or Honorable Mention at CVPR 2024 and ACL 2023. Xiang’s recent work on developing the MMMU evaluation benchmark has garnered attention beyond academia, being featured in the releases of OpenAI GPT-4o and Google Gemini. Speaker: Abstract: Large language models (LLMs) have achieved impressive progress, yet major challenges remain in enhancing their reasoning capabilities for complex tasks. In this talk, I will present our recent work on understanding and improving LLM reasoning. I will begin by discussing our efforts of understanding the reasoning, including the development of widely-used reasoning benchmarks such as MMMU and MMLU-Pro, and our studies on key factors that impact LLM reasoning performance. I will then describe our approach to improving reasoning abilities by generating large-scale synthetic reasoning data and shaping reward functions within reinforcement learning frameworks to better train reasoning models. I will conclude with a discussion on promising future directions, including how models can learn to reason more effectively through exploration and feedback. Location: CoRE 301 Committee: |
| Start Date: 03 Apr 2025; Start Time: 02:00PM - 03:30PM Title: Enhancing Deep Unrolled Models for Efficient and High-Fidelity MRI Reconstruction Bio: Speaker: Abstract: Magnetic Resonance Imaging (MRI) is a widely used imaging modality for clinical diagnostics and the planning of surgical interventions. Accelerated MRI seeks to mitigate the inherent limitation of long scanning time by reducing the amount of raw k-space data required for image reconstruction. Recently, the deep unrolled model (DUM) has demonstrated significant effectiveness and improved interpretability for MRI reconstruction, by truncating and unrolling the conventional iterative reconstruction algorithms with deep neural networks. However, the potential of DUM for MRI reconstruction has not been fully exploited. In this work, we first enhance the gradient and information flow within and across iteration stages of DUM, then we highlight the importance of using various adjacent information for accurate and memory-efficient sensitivity map estimation and improved multi-coil MRI reconstruction. Extensive experiments on several public MRI reconstruction datasets show that our method outperforms existing MRI reconstruction methods by a large margin.Publication: Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction, ECCV 2024.Fill the K-Space and Refine the Image: Prompting for Dynamic and Multi-Contrast MRI Reconstruction, MICCAIw, 2023 Location: CoRE 301 Committee: Professor Dimitris N. Metaxas (chair) Associate Professor Jingjin Yu Associate Professor Yongfeng Zhang Professor Lirong Xia |
| Start Date: 04 Apr 2025; Start Time: 10:30AM - 12:00PM Title: Multisensory Dexterity for Robotics Bio: Haozhi Qi is a final-year Ph.D. candidate in the EECS Department at UC Berkeley, advised by Prof. Yi Ma and Prof. Jitendra Malik. His research lies at the intersection of robot learning, computer vision, and tactile sensing, with the goal of developing physically intelligent, particularly dexterous, robots for unstructured environments. He received his B.S. in Mathematics and Computer Science from the Hong Kong University of Science and Technology. His work on in-hand perception was featured as the cover article in Science Robotics. He has been recognized with the Outstanding Demo Award at the NeurIPS Robot Learning Workshop and the EECS Evergreen Award for Undergraduate Researcher Mentoring. Speaker: Abstract: Human hands are essential for sensing and interacting with the physical world, allowing us to grasp and manipulate objects with ease. Replicating this dexterity in robots is the key to unlocking general-purpose robotics in unstructured environments. While modern AI has achieved breakthroughs in many domains, robot dexterity remains an unsolved challenge due to the complexity of high-dimensional control, limited real-world data, and the need for rich multisensory feedback. In this talk, I will present my work on multisensory dexterity for robotics and demonstrate how robots can achieve a broad range of dexterous manipulation capabilities. First, I will introduce how robots develop dexterous manipulation using simple sensory inputs and identify the key ingredients that enable generalizable manipulation across diverse objects, with examples in in-hand and bimanual manipulation. Building on these ingredients, I will then show how integrating rich multisensory feedback—including proprioception, vision, and tactile sensing—improves both perception and control, allowing robots to perform tasks that would be impossible with simple sensors. Finally, I will conclude with future opportunities and open challenges in scaling robotic dexterity and developing robots capable of general-purpose physical interaction. Location: CoRE 301 Committee: |
| Start Date: 07 Apr 2025; Start Time: 10:30AM - 12:00PM Title: Knowledge-Guided Machine Learning for Scientific Discovery: Challenges and Opportunities Bio: Xiaowei Jia is an Assistant Professor in the Department of Computer Science at the University of Pittsburgh. He got his Ph.D. degree from the University of Minnesota under the supervision of Prof. Vipin Kumar. His research interests include knowledge-guided machine learning and spatio-temporal data mining for real-world applications of great societal relevance. He is the recipient of the NSF CAREER Award, NASA Early Career Investigator Award, the University of Minnesota Best Dissertation Award, and multiple best paper awards from ICDM, SDM, ASONAM, and BIBE. Speaker: Abstract: Data science and machine learning (ML) models, which have found tremendous success in several commercial applications where large-scale data is available, e.g., computer vision and natural language processing, have met with limited success in scientific domains. Traditionally, physics-based models of dynamical systems are often used to study engineering and environmental systems. Despite their extensive use, these models have several well-known limitations due to incomplete or inaccurate representations of the physical processes being modeled. Given rapid data growth due to advances in sensor technologies, there is a tremendous opportunity to systematically advance modeling in these domains by using machine learning methods. However, capturing this opportunity is contingent on a paradigm shift in data-intensive scientific discovery since the "black box" use of ML often leads to serious false discoveries in scientific applications. Because the hypothesis space of scientific applications is often complex and exponentially large, an uninformed data-driven search can easily select a highly complex model that is neither generalizable nor physically interpretable, resulting in the discovery of spurious relationships, predictors, and patterns. This problem becomes worse when there is a scarcity of labeled samples, which is quite common in science and engineering domains.My work aims to build the foundations of knowledge-guided machine learning (KGML) by exploring several ways of bringing scientific knowledge and machine learning models together. In particular, we discuss gaps and opportunities in scientific discovery and show the effectiveness of KGML in multiple applications of great societal and scientific relevance. My work also has the potential to greatly advance the pace of discovery in a number of scientific and engineering disciplines where physics-based models are used, e.g., hydrology, agriculture, climate science, materials science, power engineering and biomedicine. Location: CoRE 301 Committee: |
| Start Date: 07 Apr 2025; Start Time: 02:30PM - 04:00PM Title: PROBE: Proprioceptive Obstacle Detection and Estimation while Navigating in Clutter Bio: Speaker: Abstract: In critical applications, including search-and-rescue in degraded environments, blockages can be prevalent and prevent the effective deployment of certain sensing modalities, particularly vision, due to occlusion and the constrained range of view of onboard camera sensors. To enable robots to tackle these challenges, we propose a new approach, Proprioceptive Obstacle Detection and Estimation while navigating in clutter (PROBE), which instead relies only on the robot’s proprioception to infer the presence or the absence of occluded rectangular obstacles while predicting their dimensions and poses in SE(2). The approach is a Transformer neural network that receives as input a history of applied torques and sensed whole-body movements of the robot and returns a parameterized representation of the obstacles in the environment. The effectiveness of PROBE is evaluated on simulated environments in Isaac Gym and with a real Unitree Go1 quadruped robot. The project webpage can be found at https://dhruvmetha.github.io/legged-probe/.List of publications: PROBE: Proprioceptive Obstacle Detection and Estimation while Navigating in Clutter (ICRA 2025) Location: Room 402, 4th floor, 1 Spring Street, Downtown New Brunswick Committee: Associate Professor Abdeslam Boularias Professor Kostas E. Bekris Associate Professor Jingjin Yu Professor Richard Martin |
| Start Date: 08 Apr 2025; Start Time: 10:30AM - 12:00PM Title: Rethinking AI Agents: Human-Centered Reinforcement Learning Bio: Stephanie Milani is a final-year Ph.D. candidate in the Machine Learning Department at Carnegie Mellon University. Her research focuses on building reinforcement learning agents to address human-centered and use-case-inspired challenges. Her research has been published at top machine learning and human-computer interaction venues, including ICLR, NeurIPS, and CHI, and received best paper awards at the ICML MFM-EAI and NeurIPS GenAI4Health workshops. Stephanie is a 2025 Rising Star in ML & Systems, a 2024 Future Leader in Responsible Data Science & AI, and a 2024 Rising Star in Data Science. She received the CMU Machine Learning TA award, co-organized the MineRL international competition series at NeurIPS, and received the Newman Civic Fellowship for her service to computer science education. Speaker: Abstract: : AI agents will soon be as commonplace as smartphones. These agents will make sequences of interconnected decisions that impact human lives—from serving as decision support in healthcare to shaping educational paths for millions of students. A defining challenge for the future of AI is how to build agents that can effectively operate in and adapt to these human environments.In this talk, I show how human-centered reinforcement learning offers a promising framework for addressing this challenge. First, I focus on the issue of interpretability, presenting a novel algorithm for learning transparent decision-making policies. Then, I show how human-centered design can be used to define the objectives for AI agents, exemplified through a grounded use case in mental health. Finally, recognizing that complex human domains often defy precise specification, I present our benchmark for AI agents to learn from human feedback for complex tasks. Together, this work illustrates how human-centered reinforcement learning is a valuable approach for developing AI agents that can learn from and for the people whose lives they impact. Location: CoRE 301 Committee: |
| Start Date: 10 Apr 2025; Start Time: 10:30AM - 12:00PM Title: Word Embeddings: Learning to generalize over Discrete Tokens Bio: John Blackmore is a part-time lecturer and PhD candidate in Computer Science with a concentration in Natural Language Processing (NLP). His research focuses on disambiguation of grounded language, with a critical focus on large language models (LLMs). Before devoting himself to his academic interests, John worked as a software engineer and IT architect at Educational Testing Service for 18 years, specializing in NLP Application Development. While at ETS, John contributed to several NLP research publications. He helped pioneer the use of automated scoring to grade student writing, including short answer questions, essays, and mathematical responses, as well as spontaneous speech. Speaker: Abstract: Large Language Models (LLMs) rely heavily on word embeddings as a fundamental technique. A word embedding is a vectorized representation of a word, obtained from the distributional statistics over all its occurrences in the corpus. These statistics pertain to the word(s) appearing before and after the word in the text. With the training objective of learning to predict the nearby words, the resulting matrix is used to transform each word into a vector that [many believe] captures the semantic and contextual "meaning" of the words. In this lecture, we'll discuss the theory behind word embeddings, show how they work in practice, and highlight some observations about their capabilities. Discussion is welcome and encouraged. Location: CoRE 301 Committee: |
| Start Date: 10 Apr 2025; Start Time: 09:00PM - 11:00PM Title: Deep Generative Models for Long-Horizon Decision-Making Bio: Speaker: Abstract: Model-based deep reinforcement learning (MBRL) leverages learned models of environment dynamics to facilitate more efficient planning and decision-making. Recent advancements have improved model accuracy, uncertainty estimation, and planning in latent spaces, enabling MBRL to perform competitively in complex control tasks. However, when applied to long-horizon problems, MBRL still faces critical challenges, including compounding model errors, difficulties in credit assignment under sparse rewards, and limited generalization beyond the training distribution. This dissertation addresses these challenges through innovations in both architectural design and modeling strategies, and is organized into three main parts. In the first part, we show that employing Transformer-based world models significantly improves long-horizon prediction accuracy by effectively capturing long-range temporal dependencies. However, the use of autoregressive generation introduces compounding errors over time. In the second part, we reformulate reinforcement learning as a sequence modeling problem and demonstrate that a hierarchical diffusion planner can achieve strong performance in long-horizon planning tasks. To address the limitations of diffusion planners in accurately estimating value, we further show that integrating value learning at the lower level enables the hierarchical diffusion planner to excel in both sparse-reward and dense-reward environments. Finally, in the third part, we tackle a fundamental limitation of sequence modeling approaches: their inability to generalize beyond the training distribution. As a distribution modeling technique, the diffusion planner is constrained by the data it has seen and struggles to generate valid plans when start-goal connections are absent. To overcome this, we propose a stitch-and-plan framework that enables generalization beyond the training data by composing complete trajectories from feasible sub-trajectories, effectively stitching together partial plans to form coherent solutions. Location: Virtual Committee: Professor Sungjin Ahn Professor Ruixiang Tang Professor Yongfeng Zhang |
| Start Date: 11 Apr 2025; Start Time: 10:30AM - 12:00PM Title: Cryptography-as-a-Service Bio: Aarushi Goel is an assistant professor in the computer science department at Purdue University. Prior to joining Purdue, she was a postdoctoral researcher in the Cryptography and Information Security Lab at NTT Research, mentored by Sanjam Garg. She obtained her Ph.D. from Johns Hopkins University, where she was advised by Abhishek Jain. Her research interests span broadly across cryptography and related areas of security and theoretical computer science. She was selected as a TCS and EECS rising star in 2023 and is a recipient of the Simons-Berkeley fellowship for summer 2025. Speaker: Abstract: Modern cryptography enables secure computation on private data, offering several societal benefits. For example, it empowers mutually distrustful individuals to jointly analyze their datasets while ensuring privacy. Similarly, it enables individuals to verify the integrity of private computations without compromising confidentiality. However, despite significant advancements, cryptography remains resource-intensive and has yet to see widespread adoption among end-users.My research addresses the following question: can end-users with limited resources still benefit from cryptographic advancements by outsourcing cryptographic computations to cloud servers? Moreover, can the security of such outsourced computations be guaranteed, even in the presence of untrusted servers? While cryptography can be used to ensure security when delegating “any” computation, it often introduces significant computational overheads.In this talk, I will explore whether “cryptographic” computations themselves can be securely delegated without requiring an additional, computationally expensive layer of cryptography. In other words, can we achieve efficient delegation - minimizing the burden on end-users while ensuring that it remains cost-effective? I will present new techniques that provide positive answers to these questions. Location: CoRE 301 Committee: |
| Start Date: 11 Apr 2025; Start Time: 02:00PM - 03:00PM Title: Beyond the Classroom: Careers in Computer Science Bio: Speaker: Abstract: Location: Voorhees Hall, CAC Room 105 Committee: |
| Start Date: 14 Apr 2025; Start Time: 10:30AM - 12:00PM Title: Communication, Calibration, and Grounding for Collaborative AI Agents Bio: Elias Stengel-Eskin is a Postdoctoral Research Associate at UNC Chapel Hill working with Prof. Mohit Bansal. He received his Ph.D. in Computer Science in 2023 from Johns Hopkins University under the supervision of Prof. Benjamin Van Durme, supported by an NSF Graduate Research Fellowship, and holds a Bachelor of Arts and Sciences in Cognitive Science from McGill University. His research covers a range of areas crucial to building communicative, reliable, and grounded AI agents that can handle uncertainty, ambiguity, and underspecification. His work has been featured at top conferences and journals in NLP (TACL, ACL, EMNLP, NAACL), machine learning (NeurIPS, ICLR, ICML, TMLR), computer vision (ICCV, CVPR, ECCV), and robotics (CoRL). He has also participated in writing and managing several successful grants from IARPA, NSF, and industry awards, and has served as an area chair for ACL, EMNLP, and NAACL as well as on the EMNLP 2024 organizing committee. Speaker: Abstract: To fully harness the potential of AI and ensure progress that serves all of our purposes, we will need to develop agents that can communicate and collaborate, robustly perceive their environments, and take effective and calibrated actions. In this talk, I will present work using simulated multi-agent interactions to teach models key communication skills required for teamwork, such as conveying confidence in a calibrated way and accepting or rejecting persuasion when appropriate. Moving from communication to action, I will also show how action prediction can be enriched by learning hierarchical skills over actions and code. Building on the notion of skills, I will then introduce data generation agents -- which learn to produce data for training student models based on the student's weak skills -- and show how inferred skill can inform data generation for domains such as math and code as well as multimodal settings. I will close by describing the future directions needed to develop reliable multimodal agents that interact with people in intuitive ways. Location: CoRE 301 Committee: |
| Start Date: 16 Apr 2025; Start Time: 08:30AM - 10:00AM Title: Cyber-Physical Systems for Urban Traffic Management Bio: Speaker: Abstract: My research focuses on the Cyber-Physical System(CPS) for urban traffic management, especially when abnormal situations happen. CPS is a new information paradigm that connects the physical world and the cyber world. In the physical world, we have different systems, such as agriculture systems, smart home systems, and transportation systems. Given these bodily systems, we use various sensing technologies to collect large-scale data, which will be uploaded to the cyber world for analysis, modeling, prediction, and scheduling. CPS is prone to anomalies because of the unstable interaction between the physical world and the cyber world, such as sensing noise, sensing failure, and controller failure. I'm interested in the resilient Cyber-Physical System, where we solve anomalies in the cyber world and minimize their impact on the physical world. I have developed a resilient system that solves traffic signal malfunctions, a common real-world occurrence with significant repercussions. The primary objective of this research is to mitigate the adverse effects of traffic signal malfunction, such as traffic congestion and collision, by optimizing the control of neighboring functioning signals. Additionally, I designed a Large Language Model(LLM) based framework for correcting abnormal addresses, whose occurrence leads to a significant impact on modern navigation systems. The research idea is to rewrite abnormal addresses using the strong reasoning ability of LLM and further boost LLM’s performance through objective alignment and address-centric retrieval augmented generation.Papers:MalLight: Influence-Aware Coordinated Traffic Signal Control for Traffic Signal Malfunctions, Accepted CIKM 2024AddrLLM: Address Rewriting via Large Language Model on Nationwide Logistics Data, Accpeted KDD 2024 Location: CoRE 305 Committee: Associate Professor Desheng Zhang Associate Professor Hao Wang Assistant Professor Dong Deng Assistant Professor Kangning Wang |
| Start Date: 16 Apr 2025; Start Time: 10:00AM - 11:00AM Title: Exploring Heterogeneous Graph Learning for Multi-Task Applications Bio: Speaker: Abstract: The heterogeneous graph is an important structure for modeling complex relational data and has received significant attention in recent research. However, applying heterogeneous graph learning in complex industrial scenarios presents several challenges stemming from data complexity, task complexity, and other factors. This presentation will explore and address the task complexity challenge, where industrial scenarios always involve multiple tasks within a heterogeneous graph, unlike public datasets that typically contain a single task. PapersInLINE Inner-Layer Information Exchange for Multi-task Learning on Heterogeneous Graphs under review KDD 2025JDE-Tree: A Tree-based Distribution Encoding Methods for Alleviating Information Degradation in Sub-sampling in Heterogeneous Graph Learning, Under Review CIKM 2025 Location: CoRE 301 Committee: Assistant Professor Desheng Zhang Associate Professor Hao Wang Associate Professor Yongfeng Zhang Professor Ahmed Elgammal |
| Start Date: 17 Apr 2025; Start Time: 10:30AM - 12:00PM Title: Random sampling and iteration in Python Bio: Speaker: Abstract: Description: In this teaching demo we'll talk about how to simulate randomness in Python repeatedly using NumPy and For loops. We'll talk about concepts and see them in action in Jupyter Notebook. We'll apply the skills we learn to the discussion of the Monty Hall Problem -- a counter-intuitive mathematical situation inspired by an old game show. Here too we'll both discuss it theoretically and also simulate the basic 3-door version of it in Python. Location: CoRE 301 Committee: |
| Start Date: 18 Apr 2025; Start Time: 10:00AM - 12:00PM Title: Quantum Tomography: Schur-Weyl to Pauli, 4 to 10 Bio: Nengkun Yu is a faculty member in the Department of Computer Science at Stony Brook University. He received both his Bachelor's and Ph.D. degrees from Tsinghua University. Before joining Stony Brook, he was affiliated with the University of Technology Sydney. His research interests lie in quantum learning and quantum programming. His work has been recognized with two ACM SIGPLAN Distinguished Paper Awards, at OOPSLA and PLDI, respectively. Speaker: Abstract: Determining how many copies are necessary and sufficient to identify an unknown n-qubit mixed quantum state is a fundamental problem in quantum information. In the first part of this talk, I present a tomography scheme based on the most general measurements derived from Schur-Weyl duality, achieving a sample complexity of 4^n. In the second part, I turn to Pauli measurements—widely regarded as the most experimentally accessible—and show that the required number of samples increases to 10^n. I will also discuss the applications of these tomography protocols. Location: CoRE 301 Committee: |
| Start Date: 18 Apr 2025; Start Time: 12:30PM - 02:00PM Title: Scheduling Methods for Effective Application Scaling Bio: Speaker: Abstract: In order to scale performance, modern applications make extensive use of parallelism to maximize utilization of available hardware resources. However, naive use of parallelism often fails to consider potential synchronization, variation in work complexity, heterogeneous resources, as well as caches, leaving unrealized performance on the table. This dissertation presents multiple works that examines how effective modeling and scheduling can manage such complexities and enable applications to scale more effectively against available CPU(s), storage, and memory. In the first part, we show that scheduling speculative work and weighing work based on synchronization potential can help scale a highly stateful file system checker against available CPUs and reduce overall checking runtime. In the second part, we show that dynamic migration of I/O threads across storage devices with effective performance modeling and heuristics can help scale applications against heterogenous storage and increase overall throughput performance. In the last part, we show that scheduling with effective cache latency modeling can maximize cache reuse without compromising latency, helping scale distributed VM allocation against available memory and minimize request latencies. Location: CoRE 301 Committee: Professor Sudarsun Kannan (Advisor) Professor Santosh Nagarakatte Professor Srinivas Narayana Professor Minesh Patel Professor Junaid Khalid (External Member) |
| Start Date: 21 Apr 2025; Start Time: 08:30AM - 09:30AM Title: Cyber-Physical Systems for Crowdsensing-based Indoor Mapping Bio: Speaker: Abstract: Indoor mapping is a specialized field that focuses on creating detailed representations of interior spaces. Unlike outdoor mapping, which benefits from technologies like GPS and extensive satellite imagery, indoor mapping presents unique challenges due to indoor environments confined and complex nature. Traditional methods typically rely on costly specialized devices and labor-intensive manual collection. Recent research proposes various crowdsensing-based solutions that leverage low-cost mobile devices to explore indoor environments and generate maps more efficiently. Despite its advantages, these methods still require strictly regulated data collection and device calibration procedures that hinder low-cost frequent updates. In light of this, we propose a crowdsensing-based indoor POI positioning and mapping system that leverages the courier's operation logs and order records in logistics scenarios to generate indoor maps spontaneously, without any extra effort from couriers besides their normal daily routine. Specifically, we align couriers' physical activities and digital behaviors to identify the spatiotemporal arrival event whenever they approach a POI. We creatively model the indoor POI mapping task as a classic graph realization problem. By estimating the courier's travel distances between POIs using mobile IMU inputs, we form adjacent POIs into patches, then analyze the distance geometric relations between patches to pinpoint the absolute indoor locations for POIs and generate semantic indoor POI maps accordingly.Publication: Spontaneous Indoor POI Mapping with Uncertain Courier Mobility Data, under review. Location: CoRE 305 Committee: Associate Professor Desheng Zhang Assistant Professor Dong Deng Associate Professor Yongfeng Zhang Assistant Professor Xintong Wang |
| Start Date: 21 Apr 2025; Start Time: 09:30AM - 10:30AM Title: Predictive Cyber-Physical Systems via Heterogeneous Graphs Bio: Speaker: Abstract: Cyber-physical systems (CPS) are a new information paradigm connecting the physical and cyber worlds. In CPS, we collect data from the physical world, analyze it in the cyber world, and make real-time decisions to improve the physical world. This closes the loop between the physical and cyber worlds. My work focuses on CPS via Heterogeneous Graphs to improve real-world customer experience in an industry setting. According to a recent study by Gartner, customer experience is the new battleground for business, with 81% of companies expecting to compete mostly or entirely based on customer experience by 2025.My research vision follows the CPS life cycle for customer services. I collected spatial-temporal logistics, customer, and company data from the physical world. This data is then transformed into a billion-scale heterogeneous spatial-temporal graph, and we use heterogeneous graph neural networks (HGNNs) to mine information for customer service applications. These applications, such as company key personnel detection, improve customer service through HGNNs' prediction results. The improved customer service generates new data, which is used by HGNNs to further improve prediction accuracy. In this talk, I will focus on my paper, "Paths2Pair: Meta-path Based Link Prediction in Billion-Scale Commercial Heterogeneous Graphs," which was accepted by KDD 2024. Paths2Pair can effectively discover potential company key personnel relationships from billion-scale heterogeneous graphs based on known relationships and has helped a major company identify 108,709 potential key personnel of companies.Papers:Paths2Pair: Meta-path Based Link Prediction in Billion-Scale Commercial Heterogeneous Graphs, Accepted by KDD 2024Complex-Path: Effective and Efficient Node Ranking with Paths in Billion-Scale Heterogeneous Graphs, Accepted by VLDE 2024 Location: CoRE 305 Committee: Associate Professor Desheng Zhang Assistant Professor Dong Deng Associate Professor Yongfeng Zhang Assistant Professor Sumegha Garg |
| Start Date: 21 Apr 2025; Start Time: 10:30AM - 12:00PM Title: Efficient Probabilistically Checkable Proofs from High-Dimensional Expanders Bio: Mitali Bafna is a postdoc at the Department of Mathematics at MIT, who is broadly interested in theoretical computer science. She graduated from Harvard in 2022 advised by Prof. Madhu Sudan. Her research focuses on complexity theory and algorithms, specifically combinatorial optimization, high-dimensional expanders and sum-of-squares algorithms. Her work has been awarded the Best Paper Award at STOC, 2025 and she was a Siebel Scholar (class of 2022). Speaker: Abstract: The PCP theorem, proved in the 1990s, shows how any proof can be encoded into a format that enables verification by making only a constant number of queries into the encoded proof. This landmark result in computer science has far-reaching implications for approximation algorithms and succinct verification, and PCP-based techniques are now being leveraged in blockchains like Ethereum.In this talk, I will cover some exciting progress on constructing efficient PCPs. My work builds a new set of techniques using high-dimensional expansion to construct PCPs that improve upon the state-of-the-art constructions from nearly 20 years ago. This implies that many approximation algorithms are nearly-optimal under well-believed complexity-theoretic conjectures. In the process, we also solve long-standing open problems in property testing and fault-tolerant network design. Location: CoRE 301 Committee: |
| Start Date: 21 Apr 2025; Start Time: 10:30AM - 12:30PM Title: Evolving Modern Storage Stack: I/O Abstraction, Caching and Prefetching Bio: Speaker: Abstract: With the rise of AI and data-intensive applications, the storage stack has become a critical component for performance. At the same time, new storage hardware devices, such as near-storage accelerators, are emerging. However, the storage stack has struggled to keep pace, leading to significant performance bottlenecks. This talk focuses on our works on how to evolve the modern storage stack through three key pieces: I/O abstraction, caching, and prefetching. First, we introduce a new I/O abstraction to reduce dominant I/O overheads and efficiently utilize near-storage computing capabilities. Next, we present a novel caching management solution that spans host and storage devices, optimizing memory resources. Finally, we present a cross-layered (user-level runtime and OS) prefetching mechanism for optimal performance. Location: CoRE 431 Committee: Assistant Professor Sudarsun Kannan (advisor/chair) Professor Ulrich Kremer Professor Richard Martin Assistant Professor Minesh Patel Professor Xiaosong Ma (external)Professor Xiaosong Ma (external) |
| Start Date: 22 Apr 2025; Start Time: 11:00AM - 12:30PM Title: Variational Auto Encoder (VAE): Theory and its Applications Bio: Diana Kim is a part-time lecturer at Rutgers and taught “Machine Learning Principles” to senior students in 2024 and 2025. She earned her Ph.D. in Computer Science at Rutgers in 2022, under the supervision of Prof. Elgammal at the Art and AI Lab at Rutgers, and completed a postdoc at Vision CAIR group of KAUST in Saudi Arabia in 2024. Her research is about interpreting art patterns within the latent space of various deep neural nets, using language models and art principles. Her works have been published and presented at AI conferences, including ICSC, ICCC, and AAAI 2018, 2022, and CVPR workshops 2024. Her current research focuses on creating AI vision and language systems more structured based on general domain knowledge, making them less reliant on empirical or data-driven approaches. She likes to teach students, serving as a mentor for undergraduate research internships and as a teacher for machine learning and AI classes. Before transitioning to a computer science major, her prior interests were in Electrical Engineering with a specialization in communication theory. She received an M.S. from USC (2009) and a B.S. from Ewha Womans University (South Korea, 2003). Outside academia, she worked at Samsung Electronics as a software engineer (South Korea, 2003 – 2006). Speaker: Abstract: In machine learning, computing posterior probabilities is crucial to reveal the key factors of high-dimensional data. However, finding an exact posterior solution is often intractable. Variational Auto Encoder (VAE) is a framework that enables us to find an approximated posterior, which is still rich enough to capture valuable information from data. To review the theoretical and practical aspects of VAE, in this class, we will talk about how the approximation is realized through the following components: (1) an architecture integrating the bidirectional transitions (encoder & decoder) between data and the posterior into a unified neural net, (2) optimization that targets a lower bound to tackle intractability of the original loss, and lastly (3) a parameterization trick that enables backpropagation in a VAE with a random source, which would otherwise be impossible without it. In the latter part of the class, we will focus on the practical applications of VAE. The instructor will explain how the probabilistic approach learns a continuous latent space. Through various experimental examples, we will see how the latent space is used to disentangle data factors and generate new data by incorporating the decoder block in VAE. Location: CoRE 301 Committee: |
| Start Date: 23 Apr 2025; Start Time: 01:00PM - 03:00PM Title: Circuit satisfiability: Geometric insights and polynomial threshold functions Bio: Speaker: Abstract: A fundamental question in theoretical computer science iswhether we can improve over exhaustive search for NP-completeproblems. Of these, the circuit satisfiability problem (Circuit-SAT)holds particular importance, as even modest improvements forrestricted circuit classes yield significant insights intocomputational complexity theory. In this talk, I will present somerecent progress on two frontiers of circuit satisfiability.Geometric properties of k-SAT: We develop algorithms that computegeometric properties of the set of satisfying assignments for k-CNFformulas (CNF formulas of width k). By re-analyzing the classical PPZand Schöning algorithms for k-SAT, we achieve running times comparableto state-of-the-art k-SAT solvers.Cubic threshold functions and probabilistic rank: Threshold circuits,where each gate computes a threshold function, are a powerful class ofboolean circuits with connections to neural networks and physicalcomputing models. Using the technique of probabilistic rank, we derivenew satisfiability algorithms and lower bounds for circuits withpolynomial threshold functions of degree 3 as gates.Publications: 1) https://arxiv.org/abs/2305.028502) https://www.arxiv.org/abs/2408.03465 Location: CoRE 301 Committee: Assistant Professor Karthik Srikanta Professor Michael Saks Assistant Professor Sumegha Garg Associate Professor Amélie Marian |
| Start Date: 24 Apr 2025; Start Time: 10:30AM - 12:00PM Title: Streamlining equal shares Bio: Edith Elkind is a Ginny Rometty Professor of Computer Science at Northwestern University. She obtained her PhD from Princeton in 2005, and worked in Israel, Singapore, and the UK before joining Northwestern in 2024. She works in algorithmic game theory, with a focus on algorithms for collective decision making. She is a recipient of the SIGAI Autonomous Agents Research Award and a Fellow of EurAI. She served as a chair of multiple leading conferences in AI and algorithmic game theory (including IJCAI, ACM EC, AAMAS, WINE and COMSOC), and will serve as an editor in chief of Journal of AI Research from May 2025. Speaker: Abstract: Participatory budgeting (PB) is a form of citizen participation that allows citizens to decide how public funds are spent. Through an election, citizens express their preferences over various projects (spending proposals). A voting mechanism then determines which projects will be selected. The Method of Equal Shares (MES) is a state-of-the-art proportional voting rule for participatory budgeting, which has been implemented in several European cities. Unfortunately, MES has a major drawback: it is not exhaustive, i.e., given a budget b, it may terminate after allocating b' overcome this issue, we consider Exact Equal Shares (EES) - a simpler version of MES that is known to retain its key desirable properties. We show that EES admits an efficient algorithm for iterating through virtual budgets that is guaranteed to find an optimal virtual budget; each iteration can be computed in linear time in the number of voters, and changes the outcome in a non-trivial way. Our algorithm, which we call ADD-OPT, inspires a new heuristic ADD-OPT-SKIP that shortcuts the search for an optimal virtual budget, and offers significant reduction in computation time while maintaining essentially the same level of budget utilization, as established by comprehensive experiments on real-world PB instances from Pabulib. Location: CoRE 301 Committee: |
| Start Date: 29 Apr 2025; Start Time: 10:30AM - 12:00PM Title: Quantum Hoare Logic: Towards Automatic Verification of Quantum Programs Bio: Mingsheng Ying is a Distinguished Professor at the Centre for Quantum Software and Information, University of Technology Sydney, Australia. His research interests include quantum computing, programming theory, and logics in artificial intelligence. He has authored the books Model Checking Quantum Systems: Principles and Algorithms (Cambridge University Press, 2021), Foundations of Quantum Programming (Morgan Kaufmann, 2016), and Topology in Process Calculus: Approximate Correctness and Infinite Evolution of Concurrent Programs (Springer-Verlag, 2001). He currently serves as the inaugural (Co-)Editor-in-Chief of the ACM Transactions on Quantum Computing. Speaker: Abstract: Leading technology companies, including Google, IBM, Microsoft, Intel, and Amazon, are actively advancing quantum computing by developing both hardware and software. However, programming quantum systems remains highly error-prone, as human intuition is naturally aligned with classical computation rather than quantum mechanics.In this lecture, we introduce quantum Hoare logic (QHL) — a formal framework for reasoning about the correctness of quantum programs. We will explore how QHL facilitates rigorous verification and discuss state-of-the-art tools built upon this logic. Finally, we will examine practical applications of these verification methods and their potential to enhance the reliability of quantum software. Location: CoRE 301 Committee: |
| Start Date: 29 Apr 2025; Start Time: 11:00AM - 12:30PM Title: Efficient and Scalable Management of Vector Data with Attribute Predicate Constraints Bio: Speaker: Abstract: The rise of high-dimensional vector representations, powered by advances in deep learning and representation learning, has reshaped how modern applications manage unstructured data such as text, images, and audio. Increasingly, these vector embeddings coexist with structured or semi-structured attributes—such as product metadata or image captions—creating a new class of hybrid data. While Approximate Nearest Neighbor Search (ANNS) and K-nearest neighbor graph (KNNG) construction remain foundational in vector data management, traditional indexing and retrieval techniques struggle with scalability and efficiency when handling complex queries involving diverse attribute predicates. This thesis presents a set of indexing techniques and frameworks designed to support efficient and flexible vector search under diverse predicate constraints. First, we introduce ARKGraph, a structure that compresses all possible range-aware KNN graphs, significantly reducing index size while maintaining low query latency. Second, we propose SeRF, a method that merges multiple ANNS indexes into a single compact structure, delivering efficient and consistent performance across a wide range of filtering selectivities on totally ordered attributes. Finally, we develop an attribute-agnostic search framework that integrates graph-based and inverted file indexing to support flexible queries over arbitrary predicates. Together, these contributions provide a scalable and adaptable foundation for managing hybrid datasets and enabling more expressive retrieval capabilities in real-world applications. Location: CoRE 305 Committee: Assistant Professor Dong Deng Associate Professor Desheng Zhang Associate Professor Yongfeng Zhang |
| Start Date: 30 Apr 2025; Start Time: 02:00PM - 03:30PM Title: Instruction-conditioned RL Bio: Speaker: Abstract: Reinforcement learning (RL) holds great promise in enabling autonomous systems to learn complex behaviors through interaction with their environment. However, specifying reward functions for RL is challenging due to the need for precise and often complex definitions. To address this, we propose instruction-guided reinforcement learning, where users specify tasks through instructions, eliminating the need for explicit reward functions. In the first part of this work, we consider Linear Temporal Logic (LTL) formulas to formally define instructions. Existing approaches for finding LTL-satisfying policies rely on sampling a large set of LTL instructions during training to adapt to unseen tasks at inference time, but they do not guarantee generalization to out-of-distribution LTL objectives of increased complexity. We introduce a novel approach to address this challenge, demonstrating that simple goal-conditioned RL agents can follow arbitrary LTL specifications without additional training over the LTL task space. Our approach generalizes to ω-regular expressions. Experimental results show the effectiveness of our strategy in enabling goal-conditioned RL agents to satisfy complex temporal logic task specifications zero-shot. Specifying tasks as LTLs is challenging because it requires deep knowledge of formal logic. To address this, we consider natural language for task instructions in the second part of our work. Previous studies have shown that large language model (LLM) agents can plan long-horizon tasks using a predefined set of skills from task descriptions. However, the need for prior knowledge of the required skill set limits applicability and flexibility. Our approach leverages LLMs to decompose natural language task descriptions into reusable skills, defined by LLM-generated dense reward functions and termination conditions. This facilitates effective skill policy training and chaining for task execution. To address uncertainty in the parameters used by LLMs in the generated reward and termination functions, we train parameter-conditioned skill policies that perform well across a broad spectrum of parameter values. As the impact of these parameters for one skill on the overall task becomes apparent only when subsequent skills are trained, we optimize the most suitable parameter values during the training of subsequent skills to mitigate the risk associated with incorrect parameter choices. Our experimental results show that our method is capable of generating reusable skills to solve a wide range of robot manipulation tasks.List of publicationsCompositional Reinforcement Learning from Logical Specifications. Jothimurugan et al. Learning to Reach Goals via Iterated Supervised Learning. Ghosh et al. Hindsight Experience Replay. Andrychowicz et al. Text2Reward: Automated Dense Reward Function Generation for Reinforcement Learning. Xie et al.Compositional Reinforcement Learning from Logical Specifications. Kishor et al. Location: CoRE 305 Committee: Professor Yongfeng Zhang Professor Dong Deng Professor Zihan Tan Professor He Zhu |
| Start Date: 30 Apr 2025; Start Time: 03:00PM - 05:00PM Title: Enhancing Visual Understanding with Large Foundational Models Bio: Speaker: Abstract: Large foundation models, such as vision and language models (VLMs), large language models (LLMs), and multimodal large language models (MLLMs), have demonstrated remarkable potential across a wide range of tasks, pushing us closer to achieving artificial general intelligence. However, their large model sizes pose significant efficiency challenges, particularly in vision and language tasks that require processing high-resolution images. Moreover, directly applying these models to domain-specific tasks, such as open-vocabulary object detection (OVOD), remains a non-trivial problem due to domain-specific constraints and data limitations. During my PhD program, I present a series of studies aimed at addressing efficiency and domain-specific challenges. To tackle the efficiency issue, I propose a search-based algorithm that optimizes the removal of redundant and unnecessary computations related to vision tokens. My approach accelerates MLLMs by 2× without a performance drop. Compared to existing methods, it achieves a superior efficiency-performance trade-off when the computational budgets are constrained. For domain-specific tasks, I explore strategies to leverage large foundation models to enhance domain-specific models, with a focus on OVOD. Instead of directly applying foundation models, my research investigates how these models can be utilized to address data limitations in OVOD. Specifically, I identify critical gaps in the training data that are either missing or challenging to collect. By generating synthetic data through large foundation models to fill these gaps, I demonstrate significant performance improvements in OVOD models.In a nutshell, my research improves the efficiency of large foundation models and enhances their applicability to domain-specific tasks, ultimately making them more practical and impactful in real-world scenarios. Location: CoRE 301 Committee: Professor Dimitris N. Metaxas Professor Konstantinos Michmizos Professor Ruixiang Tang Dr. Han Zhang (external) |
| Start Date: 06 May 2025; Start Time: 12:00PM - 01:30PM Title: Understanding and Enhancing Trustworthiness in Multimodal AI Bio: Speaker: Abstract: The rapid rise of generative and multimodal AI systems—such as diffusion models and Multimodal Large Language Models (MLLMs)—has unlocked new capabilities in content generation, perception, and reasoning. As these models become increasingly integrated into real-world applications, understanding and enhancing their trustworthiness—in terms of model behavior, data usage, and content safety—has become a central research challenge.In this defense, I will introduce a set of methods for interpreting and reinforcing trust in such systems. We begin with UNICORN, a framework that inverts diverse backdoor triggers in deep neural networks by formalizing trigger design spaces and decoding embedded adversarial behaviors. To address growing concerns about unauthorized data usage in generative training, I will introduce DIAGNOSIS, which injects imperceptible but learnable signal patterns into protected data and detects potential misuse by analyzing memorization traces in fine-tuned diffusion models. In addition, I present LatentTracer, a training-free method to trace the origin of images generated by latent diffusion models. By performing gradient-based latent inversion and identifying an encoder-based initialization strategy, we reveal that generated images are inherently linked to their source models—effectively acting as natural watermarks. Finally, I will present CLUE, an MLLM-as-a-Judge framework for image safety that leverages pre-trained MLLMs while overcoming the limitations of direct prompting. CLUE operates in four stages: it first objectifies subjective safety rules, then decomposes them into logically complete precondition chains; it computes image-rule relevance via debiased token probabilities, and finally applies cascaded chain-of-thought reasoning for ambiguous cases. This design enables accurate, interpretable, and updatable safety assessments—without any human annotations or model fine-tuning.Together, these contributions form coherent efforts for building interpretable, accountable, and policy-aligned trust mechanisms in modern multimodal AI systems. Location: CoRE 305 Committee: Professor Shiqing Ma Professor Dimitris Metaxas Professor Ruixiang Tang |
| Start Date: 06 May 2025; Start Time: 04:00PM - 06:00PM Title: Efficiently Manipulating Clutter via Learning and Search-Based Reasoning Bio: Speaker: Abstract: Object rearrangement is a crucial and complex problem in robotic manipulation, with applications in warehouse automation, household assistance, and industrial manufacturing. This dissertation presents novel approaches to enhance the efficiency and robustness of object manipulation planning in dynamic and cluttered environments. We introduce the Deep Interaction Prediction Network (DIPN), a learning-based model that accurately predicts object interactions, achieving over 90% accuracy in motion estimation. By integrating DIPN with Monte Carlo Tree Search (MCTS), we enable effective planning of non-prehensile actions, leading to a 100% success rate in challenging retrieval tasks. To further accelerate planning, we propose the Parallel Monte Carlo Tree Search with Batched Simulations (PMBS) framework, leveraging GPU-accelerated physics simulations to achieve a 30× speed-up. Experimental results in both simulation and real-world settings validate our approach, demonstrating state-of-the-art performance in success rates, solution quality, and computational efficiency, advancing robotic autonomy in unstructured environments. Location: Room 402, 4th floor, 1 Spring Street, Downtown New Brunswick Committee: Associate Professor Jingjin Yu Associate Professor Abdeslam Boularias Professor Kostas Bekris |
| Start Date: 07 May 2025; Start Time: 02:30PM - 04:00PM Title: PROBE: Proprioceptive Obstacle Detection and Estimation while Navigating in Clutter Bio: Speaker: Abstract: In critical applications, including search-and-rescue in degraded environments, blockages can be prevalent and prevent the effective deployment of certain sensing modalities, particularly vision, due to occlusion and the constrained range of view of onboard camera sensors. To enable robots to tackle these challenges, we propose a new approach, Proprioceptive Obstacle Detection and Estimation while navigating in clutter (PROBE), which instead relies only on the robot’s proprioception to infer the presence or the absence of occluded rectangular obstacles while predicting their dimensions and poses in SE(2). The approach is a Transformer neural network that receives as input a history of applied torques and sensed whole-body movements of the robot and returns a parameterized representation of the obstacles in the environment. The effectiveness of PROBE is evaluated on simulated environments in Isaac Gym and with a real Unitree Go1 quadruped robot. The project webpage can be found at https://dhruvmetha.github.io/legged-probe/.List of publications: PROBE: Proprioceptive Obstacle Detection and Estimation while Navigating in Clutter (ICRA 2025) Location: Room 403, 4th floor, 1 Spring Street, Downtown New Brunswick Committee: Associate Professor Abdeslam Boularias Professor Kostas E. Bekris Professor Jingjin Yu Professor Richard Martin |
| Start Date: 08 May 2025; Start Time: 01:00PM - 02:30PM Title: Full-Stack Optimization of Quantum Chemistry Simulations Bio: Speaker: Abstract: Simulating the ground state energy of molecular systems remains a central challenge in quantum chemistry and a promising application for quantum computing. Achieving accurate results on near-term quantum hardware requires careful optimization across the full computational stack—from algorithm design to algorithm compilation and hardware execution. In this work, I present two major contributions on compilation and execution, and preview my ongoing work on algorithm design. First, I introduce compiler-level techniques for optimizing the Unitary Coupled Cluster Singles and Doubles (UCCSD) ansatz, a foundational algorithm in quantum chemistry. These techniques reduce CNOT gate counts and circuit depth, significantly mitigating noise and increasing fidelity. Second, I explore quantum circuit cutting as a scalable execution strategy for large quantum circuits that exceed hardware qubit limits. By decomposing a large circuit into smaller, independently executable subcircuits and reconstructing results through classical postprocessing, this approach enables simulations beyond current hardware constraints. I further introduce a novel optimization that leverages the determinism of Clifford subcircuits to reduce experimental overhead during subcircuit tomography and memory requirements during postprocessing. Finally, I provide a preview of my ongoing work on ansatz design, focusing on balancing trainability, expressibility, noise resilience, and accuracy. These preliminary efforts aim to improve the practicality of near-term quantum simulations in quantum chemistry and lay the groundwork for further exploration in full-stack optimization.List of publications:Y. Jin et al., "Tetris: A Compilation Framework for VQA Applications in Quantum Computing," 2024 ACM/IEEE 51st Annual International Symposium on Computer Architecture (ISCA), Buenos Aires, Argentina, 2024, pp. 277-292, doi: 10.1109/ISCA59077.2024.00029.Z. Li et al., "A Case for Quantum Circuit Cutting for NISQ Applications: Impact of topology, determinism, and sparsity" arxiv: 2412.17929 Location: CoRE 229 Committee: Assistant Professor Yipeng Huang Professor Eddy Z. Zhang Professor Mario Szegedy Associate Professor Jingjin Yu |
| Start Date: 09 May 2025; Start Time: 03:00PM - 04:30PM Title: From Theory to Practice: Advancing Multi-Robot Path Planning Algorithms and Applications Bio: Speaker: Abstract: The labeled Multi-Robot Path Planning (MRPP) problem—routing multiple robots from start to goal locations without collisions—presents both fundamental algorithmic challenges and practical relevance across domains such as warehouse automation, transportation, and robotics coordination. This dissertation develops scalable MRPP solutions with theoretical guarantees and real-world applicability. On the theoretical side, it introduces algorithms with provable completeness and optimality guarantees for densely populated grid environments, as well as efficient scheduling frameworks for combined task and motion planning. On the application side, it presents effective heuristics and planning strategies for diverse settings including urban driving scenarios, robot convoy coordination, and multi-robot object delivery with constraints such as nonholonomic dynamics and tight spatial coupling. These contributions significantly advance the scalability, generality, and practicality of MRPP algorithms.The thesis defense will primarily focus on two recent lines of work. The first is the Rubik Table (RT) method, a high-throughput, near-optimal approach for MRPP on dense grid graphs, capable of coordinating tens of thousands of robots within minutes. The second is a puzzle-inspired multi-robot parking system that enables high-density vehicle storage and retrieval through combinatorial optimization, further extended to accommodate nonholonomic robot models such as Reeds-Shepp cars using motion primitives and trajectory smoothing. Both works are validated through simulations and physical experiments, demonstrating their practical impact and generalizability. Location: Room 402, 4th floor, 1 Spring Street, Downtown New Brunswick Committee: Associate Professor Jingjin Yu Professor Kostas Bekris Associate Professor Abdeslam Boularias Associate Professor Bo Yuan (External Member) |
| Start Date: 12 May 2025; Start Time: 10:30AM - 12:00PM Title: Complexity in the Era of AI and Data-Driven Computing Bio: Lance Fortnow is the inaugural Dean of the College of Computing at the Illinois Institute of Technology. At DIMACS, Fortnow represented the NEC Research Institute on the executive committee from 2000-2003 and co-chaired the 2004-2008 Special Focus on Computation and the Socio-Economic Sciences. He has authored the Computational Complexity blog since 2002 and wrote the popular science book The Golden Ticket: P, NP, and the Search for the Impossible. Speaker: Abstract: We live in a new age of computing, driven by faster distributed computation, strong optimization, data-driven algorithms, and of course dramatic advances in artificial intelligence. We’ve made dramatic progress on problems thought unsolvable a decade ago.What does this brave new world tell us about computational complexity? The P vs NP problem transforms from a barrier telling us what we cannot do to a guide to what’s possible. We are heading towards a surprising utopian computing world where we can solve many difficult problems quickly in practice while all our cryptographic protocols remain secure, and where we can make significant progress in learning in nearly every domain.We’ll give a (mostly) non-technical overview that takes a step back and rethinks complexity in light of these advances, what AI tells us about complexity, and what complexity tells us about AI. We search for not only answers, but the right questions to help us chart the future of both fields. Location: CoRE 301 Committee: |
| Start Date: 12 May 2025; Start Time: 02:00PM - 04:00PM Title: Object-Centric Manipulation: Representation and Algorithms Bio: Speaker: Abstract: In contemporary robotic manipulation research, two dominant paradigms have emerged: \textbf{action-centric} and \textbf{object-centric} approaches. Action-centric methods, exemplified by end-to-end imitation learning, emphasize direct modeling of robotic actions conditioned on sensory observations. These approaches typically operate through one or few neural networks, offering advantages in implementation simplicity and deployment efficiency. However, this paradigm demands extensive real-world robot-environment interaction data from the same type of robot for training, which is costly to collect and challenging to transfer across different robotic morphologies. In contrast, object-centric manipulation explicitly models manipulated objects and the environment through discrete representations. This framework generally follows a multistage pipeline: (1) perceptual scene understanding and object state estimation, (2) (optional) task planning based on scene analysis, (3) motion planning for action sequencing, and (4) trajectory execution via control strategies. Although architecturally more complex than action-centric alternatives due to its modular composition, the object-centric paradigm demonstrates superior cross-platform transferability and reduced simulation-to-reality transfer challenges for its separated handling of the environment and the robot. This thesis will focus on discussing object-centric manipulation methodologies, conducting a comprehensive analysis of diverse object representation schemes and their corresponding manipulation strategies across various manipulation tasks. Location: Room 402, 4th floor, 1 Spring Street, Downtown New Brunswick Committee: Professor Abdeslam Boularias (Advisor/Chair) Professor Kostas Bekris Professor Jinjin Yu |
| Start Date: 19 May 2025; Start Time: 11:00AM - 01:00PM Title: AI-Driven Correspondence Learning for Dynamic Heart Function Analysis Using MRI Bio: Speaker: Abstract: Heart disease remains a leading cause of disability and death worldwide, with various conditions adversely impacting heart function and remodeling its structures in diverse ways, leading to significant clinical consequences. Cardiac cine MRI, the gold standard for assessing heart function, is limited by inherently slow imaging speeds. Currently, most clinical cine MRI protocols are still based on 2D imaging, requiring patients to hold their breath during scans. The resulting 2D stack images often require complex post-processing, which can still lead to inaccurate and oversimplified biomarker measurements.In this defense, I will present my work on advanced AI methods for high-dimensional dynamic heart function analysis using conventional and tagged cine MRI through correspondence learning. First, I will introduce continuous spatial-temporal memory networks for 4D cardiac cine MRI segmentation. Next, I will demonstrate how neural deformable models can reconstruct 3D heart wall geometry from sparsely sampled cine MRI data. I will then explain how unsupervised learning-based image registration networks, inspired by physics, can estimate in-plane cardiac wall motion from image sequences. Finally, I will discuss volumetric neural deformable models for recovering 3D regional heart wall motion from 2D motion cues provided by tagged MRI.The defense will conclude with a vision for AI-augmented methods that have the potential to reshape the future of heart disease care. Location: CoRE 301 Committee: Professor Dimitris N. Metaxas Professor Ahmed Elgammal Associate Professor Desheng Zhang |
| Start Date: 22 May 2025; Start Time: 11:00AM - 12:00PM Title: Reachability Analysis Enables Provably Correct Controller Synthesis and Safe Exploration in Reinforcement Learning Bio: Speaker: Abstract: In this talk, we introduce how to integrate reachability analysis into Reinforcement Learning (RL) training process to enable provably correct controller synthesis and safe exploration. First, we present a verification-based learning framework VEL that synthesizes safe programmatic controllers for environments with continuous state and action spaces. VEL performs abstraction-based program verification to reason about a programmatic controller and its environment as a closed-loop system. VEL minimizes the amount of safety violation in the proof space of the system, which approximates the worst-case safety loss, using gradient-descent style optimization. Experimental results demonstrate the substantial benefits of leveraging verification feedback for synthesizing provably correct controllers. Second, we introduce VELM, a RL framework that conducts formal reachability analysis similar to VEL but for each iteration in the RL training loop with a learned symbolic environment model. An online shielding layer is then constructed to confine the RL agent's exploration solely within a state space verified as safe in the learned model, thereby bolstering the overall safety profile of the RL system. Our experimental results demonstrate that VLEM significantly reduces safety violations in comparison to existing safe learning techniques, all without compromising the RL agent's reward performance. Location: CoRE 305 Committee: Professor He Zhu Professor Srinivas Narayana Professor Santosh Nagarakatte Professor Yongfeng Zhang |
| Start Date: 04 Jun 2025; Start Time: 11:00AM - 12:30PM Title: Graph Edge Decompositions for Exploration and Visualization Bio: Speaker: Abstract: A central goal of this research is to provide tools that allow users to obtain humanly-interpretable hierarchical descriptions of any graph data. Thesetools should be accessible via a Unified Web Interface for graph analytics without being constrained by graph data size. By leveraging hierarchical graph edge partitions, our framework allows for interactive visualizations of massive graphs. Our algorithmic tools can be used to create massive data synthesis artifacts that can be fed into machine learning analytics pipelines. These include topological similarity measures for massive dataset "clustering" (e.g., Graph Cities), automatic creation of humanly-understandable summaries of social media posts (e.g., Max Flow Min Cut Views), and disentangling data communities into locally dense subgraphs with interpretable linkages between them (e.g., Community Intersection Graphs). We also illustrate an unsuspected theoretical characterization of a class of graphs derived from a partition of maximal chains in the symmetric group of permutations S_n under the Weak Bruhat Order. These graphs contain the class of visibility graphs of staircase polygons.List of publicationsCentral PapersMassive Graph Exploration[ATZ2024] Abello, J., Tangherlini, T.R., & Zhang, H. (2024). A Max Flow Min Cut View of Social Media Posts. In Proceedings of the 13th International Conference on Data Science, Technology and Applications - DATA; ISBN 978-989-758-707-8; ISSN 2184-285X, SciTePress, pages 190-201. DOI: 10.5220/0012862500003756; Best Paper Award in DATA2024, Dijon, France. [AZNHA2022] Abello, J., Zhang, H., Nakhimovich, D., Han, C., & Aanjaneya, M. (2022). Giga Graph Cities: Their Buckets, Buildings, Waves, and Fragments. IEEE Computer Graphics and Applications, 42(3), 53-64. doi: 10.1109/MCG.2022.3172650; Invited to present at IEEE VIS2023, Melbourne, Australia; Based on BigVis2021 Best Paper Award, Nicosia, Cyprus, 2021.Derived Papers [ABTZ2023] Abello, J., Broadwell, P. M., Tangherlini, T. R., & Zhang, H. (2023). Disentangling the Folklore Hairball: A Network Approach to the Characterization of a Large Folktale Corpus. Fabula, 64(1-2), 64-91. DOI: 10.1515/fabula-2023-0004[AZ2023] Abello, J., & Zhang, H. (2023). Graph Peeling Semantics. BigVis2023, CEUR Workshop Proceedings, 3379. URL: CEUR-WS.org/Vol-3379/BigVis2023_705.pdf Location: CoRE 305 Committee: Professor James Abello Associate Professor Amélie Marian Assistant Professor Karthik C.. S.. Assistant Professor Roie Levin Assistant Professor Sumegha Garg |
| Start Date: 26 Jun 2025; Start Time: 09:30AM - 11:00AM Title: Configurable Hardware-Accelerated Scientific Computing Bio: Speaker: Abstract: Parallelized accelerators are increasing in popularity and Differential Equations (DEs) are widespread in the sciences. Yet, many accelerators for DEs are not currently practical in the sciences. First, existing DE accelerators seldom demonstrate a higher order of numerical convergence that is needed to actually support existing numerical algorithms. Second, they are difficult to program, requiring low-level programming expertise. Third, most previous works are specialized stencil accelerators with narrow applicability. We demonstrate convergence for various Partial Differential Equations and Linear DEs. We show a proof-of-concept end-to-end data flow between a high-level language to configurable hardware. We successfully solve and represent the chaotic Lorenz DE as a graph with our system. Our system is over five times faster and over 100 times more energy-efficient than a CPU or GPU on a simple two-variable linear DE. Our work shows that this trifecta of high rates of convergence, friendly UI, and wide applicability are possible for DE accelerators. We thus establish that future works to advance this art should pay mind to all three aspects. Location: CoRE 305 Committee: Professor Yipeng Huang Professor Ulrich Kremer Professor Richard Martin Professor Casimir Kulikowski |
| Start Date: 05 Aug 2025; Start Time: 10:30AM - 12:00PM Title: The Evolution of Processing Units: From CPU and GPU to NPU and QPU Bio: Won Woo Ro received his B.S. degree in Electrical Engineering from Yonsei University, Seoul, Korea, in 1996, and his M.S. and Ph.D. degrees in Electrical Engineering from the University of Southern California in 1999 and 2004, respectively. He previously worked as a Research Scientist in the Department of Electrical Engineering and Computer Science at the University of California, Irvine. He is currently a Professor in the School of Electrical and Electronic Engineering at Yonsei University. Prior to joining Yonsei, he served as an Assistant Professor in the Department of Electrical and Computer Engineering at California State University, Northridge. His industry experience includes a college internship at Apple Inc. and a role as a contract software engineer at ARM Inc. His research interests span high-performance microprocessor design, GPU microarchitectures, neural network accelerators, memory hierarchy design, and quantum computing. He is serving as the General Chair for IEEE MICRO 2025 and has been inducted into the ISCA and MICRO Hall of Fame. Speaker: Abstract: In the ever-changing world of computer systems and architecture, processing units remain at the heart of technological progress. This talk takes a broad look at processing units—past, present, and future. We’ll begin by exploring the history and key role of Central Processing Units (CPUs) in advancing computing power and system design. Next, we’ll look at the rise of Graphics Processing Units (GPUs) and how they’ve transformed parallel computing and enabled today’s AI applications. As we explore the current landscape of processing technologies, the talk will highlight recent advancements that are pushing the limits of modern computing. This includes new architectures like Neural Processing Units (NPUs) and other specialized accelerators that improve performance and energy efficiency. We’ll also cover game-changing ideas like Processing-In-Memory (PIM) and in-storage computing, which move computation closer to where data lives. Finally, the talk will look ahead to what’s next—featuring breakthroughs in quantum computing and the rise of Quantum Processing Units (QPUs), which could shape the future of computing in entirely new ways. Location: CoRE 301 Committee: |
| Start Date: 12 Aug 2025; Start Time: 01:00PM - 02:30PM Title: Programmable Streaming Synchronization Scheduling Bio: Speaker: Abstract: Modern Internet applications, such as cloud gaming, autonomous vehicle sensing, and remote surgery, demand synchronization of related packet flows, known as co-flows, to ensure high-quality user experiences. Existing packet schedulers, designed for fairness or delay reduction, often fail to synchronize co-flows, leading to poor performance for these applications. This research introduces programmable abstractions that recognize co-flows, temporarily increase their priority, and use scheduling based on calendar queues to accommodate synchronized streaming on any current packet scheduler. These abstractions let packets from co-flows be temporarily paused and transmitted together or scheduled at different times to make up for delays caused by different network distances, which ensures they arrive together. A preliminary software prototype implemented as a Linux kernel module shows that flows can be synchronized with only a small amount of overhead while still being fair to cross traffic. The PSSS scheduler keeps the average time gap between synchronized co-flow packets at an average of 8 microseconds, regardless of when those packets arrive, while FIFO and FQ schedulers are very affected by delays in when packets leave the source. This work highlights the advantages of in-network streaming synchronization and lays the groundwork for effective hardware and software implementations to facilitate a variety of real-world deployments. Location: CoRE 305 Committee: Professor Srinivas Narayana Professor Rich Martin Professor Sudarsun Kannan Professor Abdeslam |
| Start Date: 26 Aug 2025; Start Time: 09:00AM - 11:00AM Title: Knowledge-Intensive and Entity-Centric Natural Language Processing Bio: Speaker: Abstract: Knowledge-intensive language processing and entity-centric language understanding are critical capabilities for modern natural language processing (NLP) systems. These abilities enable models to retrieve, integrate, and reason over large-scale external information while maintaining consistent interpretations of real-world entities. This presentation explores methods that enhance retrieval-based modeling and entity-level understanding, advancing performance on tasks such as information retrieval, retrieval-augmented generation, entity linking, and coreference resolution. In the first part, we focus on knowledge-intensive language processing. We begin with a theoretical analysis of hard negatives in the Noise Contrastive Estimation (NCE) training objective, then extend NCE to support multi-label retrieval, propose an approach to improve multi-task retrieval by encouraging task specialization, and finally introduce a retrieval-augmented generation framework that leverages implicit queries instead of human-specified ones. In the second part, we shift to entity-centric language understanding. We present a method that reformulates entity linking as an inverse open-domain question answering problem, avoiding the dilemma of predicting mentions without knowing their corresponding entities. We also propose an extremely simple yet high-performing sequence-to-sequence formulation for coreference resolution that maps input text to linearized coreference annotations. Together, these methods advance the development of NLP systems capable of deeper knowledge integration and entity-level understanding. Location: CoRE 301 Committee: Professor Karl Stratos (advisor) Associate Professor Abdeslam Boularias Assistant Professor Hao Wang |
| Start Date: 27 Aug 2025; Start Time: 01:30PM - 03:00PM Title: Physics-based Neural Deformable Models, applications to computer vision, graphics and beyond Bio: Speaker: Abstract: This thesis introduces a novel class of physics-inspired neural networks—Physics-based Neural Deformable Models(PNDMs)—that integrate traditional physics-based deformable models with modern deep learning to achieve interpretable and flexible 3D shape representations. While classical deformable models offer semantic clarity through parametric primitives, they suffer from limited geometric flexibility and dependence on handcrafted initializations. In contrast, PNDMs overcome these limitations by learning parameter functions that generalize primitive geometry, employing diffeomorphic mappings to preserve topology, and leveraging external forces for robust training.We further extend this paradigm in DeFormer, a transformer-based framework that hierarchically disentangles global and local shape deformations, and in LEPARD, which enables 3D articulated part discovery directly from 2D supervision. Finally, we demonstrate the application of our methods in photorealistic avatar reconstruction, including the LUCAS system for layered codec avatars. Together, these contributions bridge interpretable physics-based modeling with scalable neural architectures for shape abstraction, segmentation, registration, and video generation. Location: CBIM 22 Committee: Professor Dimitris Metaxas Professor Jie Gao Professor Konstantinos Michmizos Profesor Sharon Xiaolei Huang (external) |
| Start Date: 27 Aug 2025; Start Time: 03:30PM - 05:00PM Title: Towards Physics-Inspired Modeling of Multi-Agent Dynamics Bio: Speaker: Abstract: Multi-agent dynamics modeling aims to characterize the evolving states of multiple interacting agents in complex environments based on their past behaviors. It plays an important role in applications such as autonomous driving, crowd simulation and urban mobility analysis. This dissertation explores physics-inspired approaches to this problem, focusing on models which integrate physical knowledge with deep learning techniques. First, we propose a continuous-time latent variable model based on first-order neural ordinary differential equations to capture agent dynamics and interactions. To more explicitly encode physical laws, we extend this framework to second-order, acceleration-based modeling through dynamic interaction graphs. Finally, we introduce a diffusion-based generative model that captures multi-modal future behaviors conditioned on motion priors and agent interactions. Collectively, these contributions advance the modeling of multi-agent dynamics by unifying deterministic physics-based formulations, data-driven deep learning architectures, and generative models. Location: CBIM 22 Committee: Professor Dimitris Metaxas (Rutgers, chair) Professor Konstantinos Michmizos Professor Xintong Wang |
| Start Date: 29 Aug 2025; Start Time: 11:00AM - 12:00PM Title: Steering Latent Embeddings for Adaptive, Efficient, and Faithful AI Bio: Speaker: Abstract: The internal embedding space of large-scale models encodes rich semantic structure, serving as the foundation for generalization, reasoning, and adaptation. This presentation investigates how steering these latent representations can address three fundamental requirements for modern AI systems: continual adaptation, efficiency, and faithfulness. First, we introduce Contrastive Prototypical Prompting (CPP), a rehearsal-free continual learning framework that aligns and regulates class prototypes through task-specific prompts and contrastive optimization, overcoming semantic drift and prototype interference to enable long-term adaptation under strict memory constraints. Second, we study implicit in-context learning (I2CL), revealing how task adaptation can emerge purely within activation dynamics without parameter updates, and propose weakly-supervised interventions to guide these latent shifts, significantly reducing the computational and data costs of adaptation. Third, we present Visual Information Steering (VISTA), which uncovers a strong textual bias in large vision-language models and steers cross-modal embeddings to rebalance visual and textual contributions, reducing hallucination while preserving informativeness. Across these works, we demonstrate that targeted control of the embedding space—whether for stability across tasks, efficiency in adaptation, or reliability in multimodal reasoning—provides a principled pathway toward AI systems that are adaptive, efficient, and faithful in dynamic environments. Location: CBIM 22 Committee: Professor Dimitris N Metaxas (advisor) Professor Hao Wang Professor Konstantinos Michmizos |
| Start Date: 29 Aug 2025; Start Time: 02:00PM - 03:30PM Title: Towards Controllable and Efficient Diffusion Models Bio: Speaker: Abstract: Diffusion-based generative models have achieved impressive results in image and video synthesis, but their practical use is limited by high computational cost and limited controllability. In this dissertation, we address these challenges by introducing techniques to improve both the controllability and efficiency of diffusion models for image and video generation tasks. To enable fine-grained control, we develop approaches for editing a single input image under user guidance and for text-guided any-length video inpainting. To improve efficiency, we introduce a one-step image2video generation model that eliminates the need for iterative denoising, and a compact diffusion model optimized for mobile devices, enabling near real-time on-device video generation. Together, these contributions advance diffusion models toward real-world applications by enabling high-fidelity generation with greater user control and significantly reduced computational requirements. Location: CoRE 301 Committee: Prof. Dimitris Metaxas Prof. Yongfeng Zhang Prof. Konstantinos Michmizos |
| Start Date: 08 Sep 2025; Start Time: 10:30AM - 12:30PM Title: New Arenas in Hardness of Approximation Bio: Karthik C. S. is an Assistant Professor in the Department of Computer Science at Rutgers University, supported by a NSF CAREER Award, Simons Foundation Junior Faculty Fellowship, and another grant from the National Science Foundation. He received his Ph.D. in 2019 from the Weizmann Institute of Science, where he was advised by Irit Dinur, and his M.S. in 2014 from École Normale Supérieure de Lyon. He has held postdoctoral appointments at Tel Aviv University (hosted by Amir Shpilka) and New York University (hosted by Subhash Khot). He is broadly interested in complexity theory and discrete geometry with an emphasis on hardness of approximation, fine-grained complexity, and parameterized complexity. Speaker: Abstract: In this talk, I will introduce the mathematically rich area of hardness of approximation through two research programs to which I have contributed.Hardness of Approximation in P: This research program, conceived in the early 2010s by the parameterized and fine-grained complexity communities, aims to understand the approximability of problems in P such as k-Clique and k-Set Cover (for fixed k), closest pair of points, and edit distance computation through a fine-grained lens. The goal is to establish conditional lower bounds against approximation algorithms, in some cases proving that the relaxation to approximate solutions offers no significant asymptotic speedup over an exhaustive search of the solution space. I have contributed to the development of this nascent program in hardness of approximation by developing the machinery to prove some strong inapproximability results in parameterized complexity and fine-grained complexity.Inapproximability of Geometric Optimization Problems: In the 1980s and 90s, the approximation algorithms community intensively explored polynomial time approximation schemes (PTAS) for NP-hard geometric problems, culminating in the celebrated works of Arora (JACM'98) and Mitchell (SICOMP'99). Around the same time, the PCP theorem was proven, providing the backbone for proving inapproximability results for these problems. Although this spurred initial progress on the hardness of geometric problems, the field became less active after 2005. Since 2019, my work has been to revitalize these efforts, particularly in understanding the inapproximability of fundamental problems like Clustering and the Steiner Tree. Location: CoRE 301 Committee: |
| Start Date: 08 Sep 2025; Start Time: 12:00PM - 03:00PM Title: Computer Science Department Annual Student Fall Fair Bio: Speaker: Abstract: Location: Busch Student Center - MPR Committee: |
| Start Date: 11 Sep 2025; Start Time: 10:30AM - 12:00PM Title: Bayesian Deep Learning: From Reliable Neural Networks to Interpretable Foundation Models Bio: Hao Wang is currently an Assistant Professor in the Department of Computer Science at Rutgers University. Previously he was a Postdoctoral Associate at the Computer Science & Artificial Intelligence Lab (CSAIL) of MIT, working with Dina Katabi and Tommi Jaakkola. He received his PhD degree from the Hong Kong University of Science and Technology, as the sole recipient of the School of Engineering PhD Research Excellence Award in 2017. He has been a visiting researcher in the Machine Learning Department of Carnegie Mellon University. His research focuses on statistical machine learning, deep learning, and data mining, with broad applications on healthcare, network analysis, time series analysis, etc. His work on Bayesian deep learning for recommender systems has inspired hundreds of follow-up works at ICML, NIPS, ICLR, KDD, etc., becoming the most cited paper at KDD 2015 and receiving the Test of Time Award at KDD 2025. His research was recognized and supported by the Microsoft Fellowship in Asia, the Baidu Research Fellowship, the Amazon Faculty Research Award, the Microsoft AI & Society Fellowship, the NSF CAREER Award, and an NIH R01 Award. Speaker: Abstract: While perception tasks such as visual object recognition and text understanding play an important role in human intelligence, the subsequent tasks that involve inference, reasoning, and planning require an even higher level of intelligence. The past decade has seen major advances in many perception tasks using deep learning models. In terms of higher-level inference, however, probabilistic graphical models, with their ability to expressively describe properties of variables and various probabilistic relations among variables, are still more powerful and flexible. To achieve integrated intelligence that involves both perception and inference, we have been exploring along a research direction, which we call Bayesian deep learning, to tightly integrate deep learning and Bayesian models within a principled probabilistic framework. In this talk, I will present the proposed unified framework and some of our recent work on Bayesian deep learning with various applications including interpretable large language models, network analysis, and healthcare. Location: CoRE 301 Committee: |
| Start Date: 12 Sep 2025; Start Time: 12:00PM - 02:00PM Title: Beyond the Classroom: Careers in Computer Science Lecture Series Bio: Speaker: Abstract: Location: Busch Student Center -BSC Center Hall Committee: |
| Start Date: 12 Sep 2025; Start Time: 03:00PM - 05:00PM Title: Guided, Specialized, and Creative: Advancing Generative AI for Image and Video Synthesis Bio: Speaker: Abstract: Generative AI has achieved impressive progress in image and video synthesis, yet challenges remain in adapting models to new domains, personalizing outputs, ensuring temporal coherence, and fostering creativity. This dissertation aims to address these gaps. First, we present a diffusion-guided domain adaptation method enabling GANs to shift to novel domains from text prompts without target-domain data. Second, we present an approach for zero-shot personalized image generation from a single reference image using a tuning-free multimodal LLM adapter. Third, we present DirectorLLM, which decouples human motion planning from video rendering to produce temporally consistent, realistic human-centric videos. Finally, we develop a creativity-oriented diffusion framework that enhances the novelty of generated results. Together, these works advance adaptability, personalization, temporal fidelity, and creative capacity in generative modeling, a progression toward more adaptable, specialized, and creative AI systems. Location: CoRE 301 Committee: Professor Ahmed Elgammal Professor Hao Wang Professor Ruixiang Tang
|
| Start Date: 18 Sep 2025; Start Time: 03:30PM - 05:00PM Title: Exploring structured representations in Vision, Language and Abstract Reasoning Bio: Speaker: Abstract: Recent AI breakthroughs have significantly improved performance across various tasks, yet machine learning models still struggle with scene understanding and abstract reasoning—challenges that humans solve effortlessly. In this talk, we explore how structured representations can enhance AI’s ability to address these problems. First, we examine how organizing multimodal information into structured formats helps summarize scene content and address data imbalances, achieving state-of-the-art results in controllable image captioning. Next, we introduce structured approaches to abstract visual reasoning, leading to interpretable representations, enhanced generative capabilities, and improved generalization both in- and out-of-distribution. Through these advances, we highlight how structured representations can drive more efficient and interpretable AI systems. Location: CBIM 22 Committee: Professor Vladimir Pavlovic Assistant Professor Hao Wang Associate Professor Yongfeng Zhang Professor Olga Russakovsky (external)
|
| Start Date: 18 Sep 2025; Start Time: 06:00PM - 07:30PM Title: Navigating Careers Beyond the CS Degree - A Special Panel Discussion Bio: Speaker: Abstract: Location: Busch Student Center -BSC Center Hall Committee: |
| Start Date: 22 Sep 2025; Start Time: 10:30AM - 12:00PM Title: Toward Trustworthy Learning-Enabled Systems via Neurosymbolic Programming Bio: He Zhu is an Assistant Professor in the Department of Computer Science at Rutgers University. His research lies at the intersection of programming languages, formal methods, and machine learning, with a focus on building trustworthy learning-enabled systems. He is particularly interested in neurosymbolic programming, which integrates program synthesis with deep learning to enable interpretable, data-efficient, and verifiable AI. Prior to joining Rutgers, he was a research scientist at Galois, Inc., and earned his Ph.D. in Computer Science from Purdue University, where he received the Maurice H. Halstead Memorial Award. Dr. Zhu's work has been recognized with Distinguished Paper Awards at the ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI) and has been supported by the National Science Foundation and the Defense Advanced Research Projects Agency. Speaker: Abstract: Building reliable, learning-enabled systems requires bridging the gap between the interpretability and rigor of programming languages (PL) and formal methods (FM) with the flexibility of modern machine learning (ML). My research advances this integration through neurosymbolic programming (NSP), which frames learning as program synthesis: discovering programs in a domain-specific language that both optimize quantitative objectives and satisfy semantic constraints. While NSP offers a promising path toward trustworthy AI, it faces three fundamental challenges: designing suitable symbolic abstractions, navigating the combinatorial synthesis space, and ensuring correctness in systems that combine discrete programs with continuous neural components.In this talk, I will present new algorithmic frameworks that address these challenges in the context of autonomous robotic systems. First, I will describe reward-guided synthesis techniques that improve the scalability of programmatic reinforcement learning by efficiently searching large program spaces. Second, I will introduce an abstraction refinement method that automatically discovers symbolic elements for neuro-symbolic control programs, alleviating the burden of domain-specific language design. Finally, I will show how deductive reasoning and modular verification can provide correctness guarantees for neurosymbolic programs operating in continuous domains. Together, these contributions illustrate how principles from PL and FM can systematically advance NSP, enabling the construction of interpretable, sample-efficient, and verifiable learning-enabled systems. Location: CoRE 301 Committee: |
| Start Date: 26 Sep 2025; Start Time: 12:00PM - 02:00PM Title: Beyond the Classroom: Careers in Computer Science Lecture Series Bio: Speaker: Abstract: Location: Busch Student Center -BSC Center Hall Committee: |
| Start Date: 16 Oct 2025; Start Time: 10:30AM - 12:00PM Title: Approximation Algorithms: Some ancient, some new - the good, the bad and the ugly Bio: Samir Khuller received his M.S and Ph.D from Cornell University in 1989 and 1990, respectively, under the supervision of Vijay Vazirani. He spent two years as a Research Associate at the University of Maryland, before joining the Computer Science Department in 1992, where he was a Professor for 27 years. He spent several summers at the IBM T. J. Watson Research Center, and also visited the IBM Tokyo Research Lab for several weeks. From 2003 to 2008 he was the Associate Chair for Graduate Education. and he was the first Elizabeth Stevinson Iribe Chair for CS. As chair he led the development of the Brendan Iribe Center for Computer Science and Innovation, a project completed in March 2019. In March 2019, Khuller joined Northwestern University as the Peter and Adrienne Barris Chair for CS. His research interests are in graph algorithms, discrete optimization, and computational geometry. He has published about 200 journal and conference papers, and several book chapters on these topics. He was an editor for the journal Algorithmica, and International Journal on Foundations of Computer Science, problems Editor for ACM Trans. on Algorithms, and currently is a columnist for SIGACT News and Associate Editor for Networks. He has served on several program committees including SODA 1997, APPROX 1999, APPROX 2000 (chair), STOC 2003, PODS 2006, SODA 2007, APPROX 2010, ESA 2010, STOC 2013, SPAA 2017 and SODA 2021. He served on the ESA Steering Committee from 2012-2016 and chaired the 2019 MAPSP Scheduling Workshop. From 2018-2021 he will serve as the Chair of SIGACT. In 2020, he received the CRA-E Undergraduate Research Mentoring Award. He received the National Science Foundation's Career Development Award, several Dept. Teaching Awards, the Dean's Teaching Excellence Award and also a CTE-Lilly Teaching Fellowship. In 2003, he and his students were awarded the "Best newcomer paper" award for the ACM PODS Conference. He received the University of Maryland's Distinguished Scholar Teacher Award in 2007, as well as a Google Research Award and an Amazon Research Award. In 2016, he received the European Symposium on Algorithms inaugural Test of Time Award for his work with Sudipto Guha on Connected Dominating Sets. He graduated at the top of the Computer Science Class from IIT-Kanpur.
Speaker: Abstract: NP-complete problems abound in every aspect of our daily lives. One approach is to simply deploy heuristics, but for many of these we do not have any idea as to when the heuristic is effective and when it is not. Approximation algorithms have played a major role in the last three decades in developing a foundation for a better understanding of optimization techniques - greedy algorithms, algorithms based on Linear Programming relaxations have paved the way for the design of (in some cases) optimal heuristics. Are these the best ones to use in “typical” instances? Maybe, maybe not. In this talk we will focus on two specific areas - one is in the use of greedy algorithms for a basic graph problem called connected dominating set, and the other is in the development of LP based algorithms for a basic scheduling problem in the context of data center scheduling. Location: CoRE 301 Committee: |
| Start Date: 17 Oct 2025; Start Time: 12:00PM - 02:00PM Title: Beyond the Classroom: Careers in Computer Science Lecture Series Bio: Speaker: Abstract: Location: Busch Student Center -BSC Center Hall Committee: |
| Start Date: 27 Oct 2025; Start Time: 12:45PM - 02:30PM Title: Computational Aspects and Implications of Error-correcting Codes Bio: Speaker: Abstract: Error-correcting codes are large collections of strings that are pairwise far apart. Codes, by design, enjoy certain robustness guarantees, making them suitable for applications in communication, pseudorandomness, hardness of approximation, and beyond. At the same time, they give rise to a rich mathematical theory at the intersection of combinatorics, algebra, and geometry.In this talk, I will discuss various computational aspects and implications of error-correcting codes, highlighting three recent results.The first part of the talk will be devoted to permutation codes under the Ulam distance, a metric that has recently garnered attention due to applications in flash memory storage. The main highlight will be a new explicit construction of positive constant-rate codes with relative distance arbitrarily close to 1, overcoming the 1/3-distance barrier of prior constructions.The second part of the talk will concern isometric embeddings of the Hamming metric into the edit metric. Here, the main challenge is to maximize the rate, i.e., the ratio between the lengths of the input and output strings, of such embeddings. The focus will be on the first-ever constant-rate isometric embedding, along with consequences for lower bounds for problems in the edit metric.The final part of the talk will address Folded Reed-Solomon (FRS) codes — a well-studied class of codes known to achieve list-decoding capacity. The main focus will be the first fully polynomial-time algorithm running in poly(1/ε)·n·polylog(n) time for list decoding rate-R FRS codes up to radius 1-R-ε. In addition, we will see a deterministic decoder with running time f(ε)·n·polylog(n) that breaks the longstanding n^{1/ε} bound for deterministic decoding. Location: CoRE 305 Committee: Assistant Professor Karthik Srikanta Assistant Professor Roie Levin Assistant Professor Akash Kumar Sengupta Associate Professor Konstantinos Michmizos |
| Start Date: 29 Oct 2025; Start Time: 03:30PM - 05:00PM Title: Towards Robust and Generalizable Feature Representation Learning Bio: Speaker: Abstract: Feature representation learning focuses on extracting and encoding meaningful information from raw data, such as images or text. These encoded representations form the foundation for a wide range of downstream tasks, such as classification, regression, clustering, and generation. In recent years, the field has witnessed rapid progress, largely driven by advances in self-supervised learning and multimodal modeling. By leveraging large-scale, readily available datasets and designing pre-training tasks that do not rely on human annotation, researchers have been able to learn highly effective and transferable feature representations. Despite these advancements, several key challenges persist. These include: (1) designing pre-training objectives that capture the inherent structure and semantics of data, (2) ensuring learned features generalize across diverse downstream tasks, and (3) learning effectively from noisy, weakly aligned data.This research focuses on addressing these issues. Specifically, the dissertation investigates feature representation learning across multiple modalities, including human skeleton data, visual data (images and videos), and text. It explores the full spectrum of representation learning challenges, covering the design of pre-training tasks, the formulation of representation structures, and the mitigation of noise in large-scale data. To this end, the dissertation presents: (1) a hierarchical encoder combined with a pretext-based self-supervised framework for modeling structured skeleton data; (2) a codebook-based representation strategy that improves alignment between image and text modalities and addresses semantic mismatches during pretraining; and (3) an automatic data filtering system that leverages language models and multi-pathway alignment to filter noisy supervision in video-language data. Together, these contributions offer a unified perspective on learning robust, transferable, and interpretable feature representations across diverse domains and challenges. Location: CBIM 22 Committee: Prof. Dimitris N. Metaxas (advisor/chair) Prof. Hao Wang Prof. Konstantinos Michmizos |
| Start Date: 31 Oct 2025; Start Time: 11:00AM - 12:30PM Title: Adventures in High-Dimensional Quantum Codes Bio: Speaker: Abstract: Quantum computers have steadily improved over the last decade, but developing fault-tolerant quantum computing (FTQC) techniques, required for useful, universal computation remains an ongoing effort. Key elements of FTQC such as error-correcting codes are supported by a rich bed of stabilizer simulation software such as Stim and CHP, which are essential for numerically characterizing these protocols at realistic scales. Namely, tasks like benchmarking codes' practical thresholds under different error decoding strategies can only be studied via simulation. Recently, experimental groups have built nascent high-dimensional quantum hardware, known as qudits, which have a myriad of attractive properties for algorithms and FTQC. Despite the stage being set for early high-dimensional FTQC, there are no widely available qudit stabilizer simulators. We introduce Sdim, the first open-source realization of such a simulator for all dimensions. We demonstrate its correctness against existing state vector simulations and benchmark its performance in evaluating and sampling quantum circuits. Qudits, while being built out of the same devices as qubits, have access to states on a larger Hilbert space and a potentially richer structure for error correction, but neither current simulation nor experiment have supported a clear affirmative or negative. This simulator is the essential computational infrastructure by which we may systematically answer these questions, and the answers will guide error-correction efforts down the most fruitful paths towards FTQC. Location: CoRE 301 Committee: Prof. Yipeng Huang Prof. Mario Szegedy Prof. Emina Soljanin Prof. Tomasz Imielinski |
| Start Date: 31 Oct 2025; Start Time: 12:00PM - 02:00PM Title: Beyond the Classroom: Careers in Computer Science Lecture Series Bio: Speaker: Abstract: Location: Busch Student Center -BSC Center Hall Committee: |
| Start Date: 04 Nov 2025; Start Time: 10:30AM - 12:00PM Title: Generalizability and Machine Learning in Biomedicine Bio: Speaker: Abstract: There are often performance differences between intra-dataset and cross-dataset tests in machine learning (ML) modeling of biomedical data, as well as those between applying ML to all samples and a subset of them (e.g., all versus older patients). However, reducing these differences may reduce ML performances. It is thus a challenging dilemma to develop models that excel in intra-dataset testing and are generalizable to subset-sample or cross-dataset testing. Therefore, we propose a multi-criteria framework to 1). Improve ML fairness in classifying multicategory cause of deaths in cancer patients; and 2). Understand and improve performance and generalizability of ML in intra-dataset and cross-dataset testing. Among the colorectal cancer patients (n=515) of various age, sex, and racial groups in the TCGA data, all ML models exhibited biases for these sociodemographic groups. Methods to optimize model performance, including testing the model on merged groups and others, show the potential to reduce disparities in model performance for different groups. Importantly, both robust Analysis of Variance (ANOVA) and Kruskal–Wallis tests consistently identified differentially expressed genes as one of the most influential factors in both cancer types. The proposed multi-criteria framework successfully identified the model that achieved both the best cross-dataset performance and similar intra-dataset performance. In summary, generalizing ML performance is challenging, as evidenced by ML biases in classifying a subset of samples and lower performance in cross-dataset testing. We seem able to develop methods to improve ML fairness and generalizable ML performance. Specifically, ML performance distributions significantly deviated from normality, which motivates using both robust parametric and non-parametric statistical tests. We also quantified and provided possible exploitability on the factors associated with cross-dataset performances and generalizability of ML models in two cancer types. A multi-criteria framework was developed and validated to identify the models that are accurate and consistently robust cross datasets Location: CoRE 301 Committee: |
| Start Date: 06 Nov 2025; Start Time: 10:30AM - 12:00PM Title: Chasing the Constant: Bridging Theory and Practice in Privacy-Preserving Machine Learning Bio: Speaker: Abstract: At the heart of the first large-scale deployments of Google Gboard’s private next-word prediction and Apple’s private federated learning framework lies a simple and perhaps the most fundamental primitive: differentially private counting in the continual release model. This primitive serves as a subroutine not only in federated learning, but also in a wide range of applications, including histogram estimation, non-interactive local learning, graph statistics, stochastic convex optimization, and matrix analysis.In this talk, I will present my recent work establishing deep connections between private continual counting and a concept in operator algebra (factorization norms), which not only advances the foundations of privacy-preserving learning but also improves more than three decades-old results in operator algebra. I will also highlight my results in private graph analysis, where my work resolved an open problem in differential privacy, the first efficient and exact sampling algorithm from a logconcave distribution defined over non-convex sets, and improved results in discrepancy of shortest paths. I will conclude with my future plans. My goal is to design next-generation privacy-preserving deep learning algorithms that adapt to time-varying data sensitivity and system constraints, and to extend these methods into decentralized AI/ML solutions that reduce reliance on large data centers. This vision aims to make privacy-preserving machine learning both theoretically rigorous and practically deployable, paving the way for scalable, sustainable, responsible, and inclusive AI systems. Location: CoRE 301 Committee: |
| Start Date: 07 Nov 2025; Start Time: 05:00PM - 06:30PM Title: Antithetic Noise in Diffusion Models Bio: Speaker: Abstract: We initiate a systematic study of antithetic initial noise in diffusion models. Across unconditional models trained on diverse datasets, text-conditioned latent diffusion models, and diffusion posterior samplers, we find that pairing each initial noise with its negation consistently yields strongly negatively correlated samples. To explain this phenomenon, we combine experiments and theoretical analysis, leading to a symmetry conjecture that the learned score function is approximately affine antisymmetric (odd symmetry up to a constant shift), and provide evidence supporting it. Leveraging this negative correlation, we enable two applications: (1) enhancing image diversity in models like Stable Diffusion without quality loss, and (2) sharpening uncertainty quantification (e.g., up to 90% narrower confidence intervals) when estimating downstream statistics. Building on these gains, we extend the two-point pairing to a randomized quasi-Monte Carlo estimator, which further improves estimation accuracy. Our framework is training-free, model-agnostic, and adds no runtime overhead.Link: https://arxiv.org/pdf/2506.06185 Location: CoRE 305 Committee: Professor Peng Zhang Professor Ruixiang Tang Professor Hao Wang Professor Yipeng Huang |
| Start Date: 14 Nov 2025; Start Time: 12:00PM - 02:00PM Title: Beyond the Classroom: Careers in Computer Science Lecture Series Bio: Speaker: Abstract: Location: Busch Student Center -BSC Center Hall Committee: |
| Start Date: 17 Nov 2025; Start Time: 10:00AM - 12:00PM Title: Almost Sharp Bounds on Weaver's Discrepancy for Gaussian Vectors Across Different Regimes Bio: Speaker: Abstract: Weaver's discrepancy measures the minimum operator norm of signed sums of rank-one matrices and plays an important role in discrepancy theory and applications such as the Kadison-Singer problem, graph sparsification, and experimental design. While the worst-case bound is well understood following the Marcus-Spielman-Srivastava resolution of the Kadison-Singer conjecture (Annals of Mathematics, 2015), the average-case behavior remains less explored.In this talk, we will study Weaver's discrepancy under a probabilistic model where the input vectors are drawn independently from the standard Gaussian distribution in $\mathbb{R}^d$.We will provide almost tight bounds on the discrepancy across a wide range of dimensions $d$ relative to the number of vectors $n$. In the extremely subcritical regime $d = o(n^{1/4})$, we establish almost matching upper and lower bounds that are exponentially smaller than the Marcus-Spielman-Srivastava bound, and our lower bounds are valid in the entire subcritical regime $d = o(n^{1/2})$.In the supercritical regime $d = \omega (n^{1/2})$, we show that the discrepancy is $\Omega(\sqrt{dn})$ with high probability, matching known upper bounds and demonstrating that the Marcus-Spielman-Srivastava bound is tight even for Gaussian random inputs. Our results give a nearly complete characterization of the average-case discrepancy for Gaussian vectors and clarify the critical threshold that separates different asymptotic regimes.Related Publications:Ziyi Cai, Qing Chen & Peng Zhang. (2025). WEAVER'S DISCREPANCY FOR GAUSSIAN RANDOM VECTORS. SIAM Journal on Discrete Mathematics, 39(3), 1418-1447. https://doi.org/10.1137/24M1678878Qing Chen, Yunwei Ren & Peng Zhang. ALMOST SHARP BOUND ON WEAVER'S DISCREPANCY FOR GAUSSIAN VECTORS ACROSS DIFFERENT REGIMES. In submission. Location: CoRE 305 Committee: Assistant Professor Peng Zhang Professor Jie Gao Assistant Professor Kangning Wang Assistant Professor Hao Wang |
| Start Date: 18 Nov 2025; Start Time: 10:30AM - 12:00PM Title: The Many Facets of Monte Carlo: Quantum-Inspired Averaging and Antithetic Diffusion Sampling Bio: Speaker: Abstract: Monte Carlo methods play a crucial role in statistics, computer science, and physics. I will present two such facets. The first is a quantum-inspired averaging process on graphs, a simplified abstraction of Google’s quantum sampling experiment, where local random interactions drive convergence toward equilibrium. We analyze its mixing time and establish universal lower bounds on how fast such distributed sampling dynamics can approach uniformity. The second comes from diffusion-based generative models, where pairing each Gaussian noise with its negation induces strong negative correlation. This simple antithetic design sharply reduces sampling variance and improves reliability in downstream estimation tasks. Location: CoRE 301 Committee: |
| Start Date: 01 Dec 2025; Start Time: 03:30PM - 05:30PM Title: Controllable and Efficient Generative Models: Methods and Applications in Medical Imaging Bio: Speaker: Abstract: Deep generative models have revolutionized visual computing, achieving unprecedented realism in synthesizing images, videos, and 3D scenes. However, as the quality of generation matures, the critical challenge shifts towards achieving precise, user-driven control and practical, efficient deployment, especially in high-stakes domains like medical imaging. This dissertation addresses two fundamental obstacles hindering this transition. The first is the "controlled generation" in emerging generative architectures, such as discrete diffusion and next-scale visual autoregressive models, whose non-invertible sampling mechanisms prevent the recovery of latent codes necessary for high-fidelity editing. The second is the "practicality gap" in clinical applications, where a confluence of data scarcity (few-shot learning), data heterogeneity (non-IID distributions), stringent privacy constraints (requiring Federated Learning), and communication bottlenecks impede the collaborative development of robust AI models.This dissertation presents a cohesive body of work that bridges these gaps, progressing from foundational algorithmic innovations to their application in real-world medical imaging. The primary contributions are fourfold: (1) DICE, a pioneering inversion framework that enables, for the first time, controllable editing for discrete diffusion and masked generative models; (2) VARIN, the first noise inversion-based editing technique for next-scale visual autoregressive models, which introduces a novel pseudo-inverse for the argmax operator; (3) DMCVR, a morphology-guided diffusion model that solves the clinical problem of 3D cardiac volume reconstruction from sparse MRI by leveraging explicit anatomical conditioning; and (4) a comprehensive suite of Federated Learning frameworks that integrate few-shot learning, dual knowledge distillation, and parameter-efficient fine-tuning (LoRA) to enable robust, private, and communication-efficient medical image analysis on decentralized data. Collectively, this research provides a validated blueprint for building the next generation of visual computing systems that are not only powerful but also controllable, efficient, and trustworthy. Location: CoRE 301 Committee: Prof. Dimitris N. Metaxas (Chair) Prof. Konstantinos Michmizos Prof. Hongyi Wang |
| Start Date: 11 Dec 2025; Start Time: 10:00AM - 11:30AM Title: Metric Distortion of Matching: A Stochastic Perspective Bio: Speaker: Abstract: Bipartite matching is a fundamental problem with a wide range of applications, in many of which only ordinal preferences, rather than the underlying cardinal utilities, are available. The metric distortion framework was established to measure the efficiency loss when only ordinal information is used. However, determining the optimal metric distortion for bipartite matching has proved to be notoriously difficult.In this talk, I will present some recent progress on metric distortion of matching under stochastic models. Average-case distortion on a uniform line: We show that when agents and items are drawn i.i.d. from a uniform distribution on a one-dimensional segment, the simple matching rule of random serial dictatorship achieves exponentially better distortion than the best-known distortion in the worst-case model. Semi-random model: When points in an arbitrary metric space are randomly partitioned into two sides, we propose a new simple matching rule with distortion better than the best-known distortion in the worst-case model.Related Publication(s): Ziyi Cai, Qing Chen, Kangning Wang, and Peng Zhang. Metric Distortion of Matching: A Stochastic Perspective. In submission. Location: CoRE 305 Committee: Assistant Professor Peng Zhang Assistant Professor Kangning Wang Assistant Professor Xintong Wang Professor Eddy Z. Zhang |
Upcoming Events
| 07 Nov 2025; - 05:00PM - 06:30PM Antithetic Noise in Diffusion Models |
| 14 Nov 2025; - 12:00PM - 02:00PM Beyond the Classroom: Careers in Computer Science Lecture Series |
| 17 Nov 2025; - 10:00AM - 12:00PM Almost Sharp Bounds on Weaver's Discrepancy for Gaussian Vectors Across Different Regimes |