Abstract: Generative image models trained on massive datasets encode the statistics of our visual world. An off-the-shelf diffusion or flow-matching model can therefore serve as a general-purpose image prior, providing information about images across a wide range of tasks. Harnessing these capabilities, however, requires combining the model at inference time with additional signals or constraints unknown during training. This thesis will develop training-free methods for inference under a generative denoising prior. I first show how sampling from a trained denoiser can be formulated as an optimization problem, which combined with a constraint, can solve tasks ranging from conditional generation and weakly supervised segmentation to combinatorial optimization. I then demonstrate how the inherent properties of denoisers can accelerate inference under such constraints. Replacing gradient descent with an inexact Newton update, based on the symmetry of the denoiser's Jacobian, substantially reduces inference costs without tradeoffs. I also explore a middle ground between model adaptation and fully training-free inference by using the denoiser's robust internal representations to learn constraints from limited labeled data. These methods are applied to gigapixel image domains such as digital histopathology and remote sensing, where generative models can only be trained at a patch scale. To synthesize arbitrarily large images at resolutions unseen during training, I introduce inference-time algorithms that enforce consistency across spatially overlapping patches and image scales. Finally, I propose repurposing inference-time algorithms from sampling tools, to mechanisms for understanding the prior learned by a denoising model. Extending the previously developed techniques, I analyze the Jacobians of generative denoisers, where preliminary results indicate that Jacobian spectra correlates with generative quality. This motivates a Jacobian-spectrum regularization as a way to improve model performance using insights derived from inference-time algorithms.

Speaker: Alexandros Graikos

Location: NCS 220


Abstract: In high-dimensional data spaces, vast empty regions often exist where no known data points are present. These empty spaces are not merely gaps but hold untapped potential for discovering novel configurations, optimizing parameters, and improving decision-making processes. However, traditional exploration techniques struggle to identify and leverage these regions due to the curse of dimensionality. To address this, we introduce the Empty Space Search Algorithm (ESA), a scalable, physics-inspired method that systematically identifies and explores these uncharted voids. ESA operates by modeling the data space as a dynamic system, using a repulsion-attraction mechanism to locate optimal empty space configurations (ESCs) without requiring exhaustive search. Building upon ESA, we present GapMiner, a visual analytics system that integrates human-in-the-loop AI to iteratively refine and validate ESCs. GapMiner combines parallel coordinate visualization, interactive optimization, and deep learning-based predictive modeling to enhance the efficiency of empty space exploration. This methodology has broad applications, including accelerating convergence in evolutionary algorithms through a more diverse initial population, optimizing adversarial learning strategies, and discovering novel parameter configurations in reinforcement learning. Our approach demonstrates that empty space is not just an absence of data but a frontier for new possibilities in high-dimensional problem-solving.
Bio: Xinyu Zhang received his B.E. in Computer Science from Shandong University, Taishan College, in 2019. He is currently a final-year Ph.D. candidate in the Department of Computer Science at Stony Brook University, advised by Prof. Klaus Mueller. His research focuses on multivariate data analysis, scientific visualization, and reinforcement learning. He has published multiple papers in top-tier journals and conferences, including IEEE TVCG and NeurIPS.
*this seminar will be held in person (food provided on a first come, first serve basis), and online (zoom link below)!
Topic: IACS Student Seminar Speaker: Xinyu Zhang
Time: Feb 26, 2025 12:00 PM Eastern Time (US and Canada)
Join Zoom Meeting
https://stonybrook.zoom.us/j/91848218975?pwd=lfITFa61GaXZ2Wsa1B1OnbLQMmXvOE.1

Meeting ID: 918 4821 8975
Passcode: 027337

We will demonstrate how to make a personalized AI partner that can do certain tasks for you. This step by step process can be used to develop your own companion.

Understand what is a Gemini Gems and how it can be personalized for your custom needs. We will show you how to capture the tasks that you would like to complete and give you higher quality responses that will ensure your communication and tasks are completed with ease.

Register here.

TITLE: Towards a Theory of Encode/Decoder Architectures by Andrej Risteski of CMU

ABSTRACT: A common choice of architecture in representation learning (i.e., learning a good embedding of the data) is an encoder/decoder architecture, which tries to map a part of the input into a good latent representation (via an encoder), and predict the remaining part of the input (via a decoder). Two common examples are universal machine translation: where one tries to learn to translate between any pair of a set of languages via a common latent language, given paired up corpora for only a part of the pairs; and contextual encoders -- where one tries to predict a part of the image, given the rest of the image.
 
We will give a framework for analyzing the sample complexity of such architectures -- i.e., how many pairs of languages do we need to have paired up corpora for? How many image prediction tasks do we have to solve to get a good representation?
The Natural Language Processing Reading Group at Stony Brook University meets weekly to discuss recent research papers in NLP and related fields.
Join the Google Group here.

Title: J-Space and Mechanistic Interpretability

Abstract, J-Space: The rise of the term mechanistic interpretability has accompanied increasing interest in understanding neural models -- particularly language models. However, this jargon has also led to a fair amount of confusion. So, what does it mean to be mechanistic? We describe four uses of the term in interpretability research. The most narrow technical definition requires a claim of causality, while a broader technical definition allows for any exploration of a model's internals. However, the term also has a narrow cultural definition describing a cultural movement. To understand this semantic drift, we present a history of the NLP interpretability community and the formation of the separate, parallel mechanistic interpretability community. Finally, we discuss the broad cultural definition -- encompassing the entire field of interpretability -- and why the traditional NLP interpretability community has come to embrace it. We argue that the polysemy of mechanistic is the product of a critical divide within the interpretability community.

Abstract, Mechanistic Interpretability: If the mind is an ocean, we spend our lives floating at the surface. Beneath us, an enormous amount of processing takes place without our knowledge: our visual systems parsing the contours of a face, our motor circuits maintaining our posture. At any given moment, only a small fraction of this neural activity is accessible to us. Yet it is this privileged sliver of activity that we rely on to reason deliberately: to plan what ingredients to buy for a recipe, or to puzzle out why an engine won't start. Such thoughts can be articulated out loud, deliberately held in mind, and brought to bear on whatever task the moment demands. This distinction, between our accessible thoughts and our unconscious processing, is perhaps the most striking feature of human cognition. In this paper, we present evidence that an analogous functional distinction has emerged in modern AI models. Specifically, we observe that language models maintain a privileged set of internal representations, available for report, modulation, and flexible internal reasoning, atop a much larger volume of automatic processing. We identify these representations using a new interpretability technique, which surfaces the concepts a model is poised to verbalize at any point in its processing. Measuring and intervening on these representations provides us a window into a model's thought processes, uncovering internal reasoning and reactions that do not appear in its output.

Location: NCS 220

Zoom Link: https://stonybrook.zoom.us/j/92942833294?pwd=ysKMaQx7Hp0lkq5nOb3zJ9ivNZVvLv.1&jst=2
The IEEE/CVF Conference on Computer Vision and Pattern Recognition 2025 will be held from June 11th to June 15th, 2025, at the Music City Center, Nashville, TN. The IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR) is the premier annual computer vision event comprising the main conference and several co-located workshops and short courses. With its high quality and low cost, it provides an exceptional value for students, academics and industry researchers. Register here.
Fall 2026, Wednesdays 2 to 3:20 pm, NCS 220 and Zoom link to be announced soon.

The seminar will be jointly taught by Prof. Dimitris Samaras (samaras@cs.stonybrook.edu).

The overall purpose of this seminar is to bring together people with interests in Computer Vision theory and techniques and to examine current research issues. This course will be appropriate for people who already took a Computer Vision graduate course or already had research experience in Computer Vision.

To enroll in this course, you must either: (1) be in the Ph.D. program or (2) receive permission from the instructors.

Each seminar will consist of multiple short talks (around 15 minutes) by multiple students. Students can register for 1 credit for CSE656. Registered students must attend and present a minimum of 2 talks. Registered students must attend in person. Up to 3 absences will be excused. Everyone else is welcome!

Abstract: The landscape of machine learning evolves rapidly and the complexity of the networks and their architectures defies easy comprehension. AI is touted as the next scientific revolution by allowing the processing and pattern-finding in increasingly massive data sets. One potential end results could be AI enhanced measurement technologies, but what does that mean? This talk will give examples of how classical tools indicate the technical obstacles to this vision in terms of understanding training processes, model comparisons, and feature embeddings. While the results in this talk are largely empirical, they point to interesting directions for (infomation?) theoretical investigation.

Bio: Anand D. Sarwate is an Associate Professor in the Electrical and Computer Engineering Department at Rutgers, The State University of New Jersey. He received B.S. degrees in math and electrical engineering from MIT and a Ph.D. in electrical engineering from UC Berkeley. Prior to joining Rutgers he was a Research Assistant Professor at TTI-Chicago and a postdoc at the ITA Center at UC San Diego. His research interests include information theory, machine learning, signal processing, optimization, and privacy and security.
Location: Light Engineering 250
The SUNY Office of Research, Innovation & Economic Development (ORIED) is hosting a webinar, Pathways to Innovation: Exclusive STEM Opportunities for Students at Premier Labs, with the Air Force Research Laboratory (AFRL), the Griffiss Institute and Brookhaven National Laboratory (BNL).

Please join us on October 30 from 12:30 - 2:00 pm to learn more about the labs and the wide variety of research, education, and workforce development programs they offer.

Register here: https://rfsuny.zoom.us/webinar/register/WN_fjWNU9l8Sr6WO_M3AoZ-Rw?mc_cid=50c2045945&mc_eid=357e15f9df#/registration