Abstract: Many unresolved legal questions over LLMs and copyright center on memorization: whether specific training data have been encoded in the model's weights during training, and whether those memorized data can be extracted in the model's outputs. While many believe that LLMs do not memorize much of their training data, recent work shows that substantial amounts of copyrighted text can be extracted from open-weight models. However, it remains an open question if similar extraction is feasible for production LLMs, given the safety measures these systems implement. We investigate this question using a two-phase procedure: (1) an initial probe to test for extraction feasibility, which sometimes uses a Best-of-N (BoN) jailbreak, followed by (2) iterative continuation prompts to attempt to extract the book. We evaluate our procedure on four production LLMs -- Claude 3.7 Sonnet, GPT-4.1, Gemini 2.5 Pro, and Grok 3 -- and we measure extraction success with a score computed from a block-based approximation of longest common substring (nv-recall). With different per-LLM experimental configurations, we were able to extract varying amounts of text. For the Phase 1 probe, it was unnecessary to jailbreak Gemini 2.5 Pro and Grok 3 to extract text (e.g, nv-recall of 76.8% and 70.3%, respectively, for Harry Potter and the Sorcerer's Stone), while it was necessary for Claude 3.7 Sonnet and GPT-4.1. In some cases, jailbroken Claude 3.7 Sonnet outputs entire books near-verbatim (e.g., nv-recall=95.8%). GPT-4.1 requires significantly more BoN attempts (e.g., 20X), and eventually refuses to continue (e.g., nv-recall=4.0%). Taken together, our work highlights that, even with model- and system-level safeguards, extraction of (in-copyright) training data remains a risk for production LLMs.

Speaker: Xinyue

Location: CS2311

Abstract: Pretraining vision encoders with self-supervision (SSL) leads to stronger representations that excel across diverse downstream tasks. One of the key factors enabling self-supervision is extracting multiple views of the same scene to formulate either: 1) View-invariant pretraining (DINO, SimCLR, iBOT), where the objective is predicting the same representation for different views of the scene; or 2) Cross-view pretraining (cross-view Masked Autoencoders), where the objective is predicting missing parts of one view using other views. For extracting multiple views, view-invariant methods rely on a combination of handcrafted augmentations (random cropping, color jittering, gaussian blur, etc.) of the same image, whereas cross-view pretraining methods rely on image cropping or video frames. In this work, we present methods to effectively incorporate synthetic views from diffusion models into SSL training.
For view-invariant pretraining, we introduce Gen-SIS, a method that leverages the ability of diffusion models to generate interpolated images through interpolation in conditioning space. We introduce a disentanglement pretext task: disentangling two source images from an interpolated synthetic image. This disentanglement task, in addition to vanilla single-source generative augmentation for view extraction, improves visual pretraining of various view-invariant methods (DINO, SimCLR, iBOT).
For cross-view pretraining, we introduce CDG-MAE, a novel cross-view masked autoencoder (MAE) based method that uses diverse synthetic views generated from static images via an image-conditioned diffusion model to learn dense correspondences. We present a quantitative method to evaluate the local and global consistency of the generated views to choose the right diffusion model for cross-view pretraining. These generated views exhibit substantial changes in pose and perspective, providing a rich training signal that overcomes the limitations of video (expensive) and crop-based (less variation) methods. CDG-MAE substantially narrows the gap to video-based MAE methods on video label propagation tasks while maintaining the data advantages of image-only MAEs.

Speaker: Varun Belagali

Location: NCS 120
Zoom: https://stonybrook.zoom.us/j/93647452432?pwd=hZaX7LXCAD8KPHWYE1Afw2sDI3owpv.1
Join Zoom Meeting
https://stonybrook.zoom.us/j/91945227869?pwd=emhoZDFWVTV0MVdPWW5uVk43MjQzUT09

Meeting ID: 919 4522 7869
Passcode: 452304
One tap mobile
+16468769923,,91945227869# US (New York)
+13126266799,,91945227869# US (Chicago)

Dial by your location
        +1 646 876 9923 US (New York)
        +1 312 626 6799 US (Chicago)
        +1 301 715 8592 US (Germantown)
        +1 669 900 6833 US (San Jose)
        +1 253 215 8782 US (Tacoma)
        +1 346 248 7799 US (Houston)
        +1 408 638 0968 US (San Jose)
Meeting ID: 919 4522 7869
Find your local number: https://stonybrook.zoom.us/u/aCvAYWkRg
  
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
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 to attend. The seminar will be taught by Prof. Chao Chen, chao.chen.1@stonybrook.edu.
Abstract: Many foundation models for digital pathology have been released recently. Benchmarking available methods then becomes paramount to get a clearer view of the research landscape. For this reason, we introduce THUNDER, a tile-level benchmark for digital pathology foundation models, allowing for efficient comparison of many models on diverse datasets with a series of downstream tasks, studying their feature spaces and assessing the robustness and uncertainty of predictions informed by their embeddings. Such foundation models are often used as feature extractors and combined with Multiple Instance Learning (MIL) aggregators at downstream time. Such aggregation must be efficient and reliable. We will focus on two specific examples of this: (I) HistAug, a fast and efficient generative model for controllable augmentations in the latent space of foundation models to perform data augmentation for MIL, and (ii) CAR-MIL, a method based on counterfactual attention regularisation to improve the reliability of attention maps of MIL methods.

Short-bio: Pierre Marza is a Postdoctoral Researcher at CentraleSupelec in the Biomathematics team of the MICS lab, studying Computer Vision and Deep Learning for Medical Imaging, with a focus on Digital Pathology. Prior to this, he was a PhD student at INSA Lyon, in the LIRIS and CITI labs, advised by Christian Wolf, and co-advised by Laetita Matignon and Olivier Simonin. He studied Visual Navigation, Embodied AI, Spatial Reasoning, more specifically how to learn to represent 3D space, generalize to new environments and master diverse tasks from light supervision.

Location: NCS 220

Zoom: https://stonybrook.zoom.us/j/94798224254?pwd=CFraer25qnpORbJ14aAVHRwaSJOjJM.1
Hidden Biases. Ethical Issues in NLP, and What to Do about Them presented by Dirk Hovy of Bocconi University

ABSTRACT: Through language, we fundamentally express who we are as humans. This property makes text a fantastic resource for research into the complexity of the human mind, from social sciences to humanities. However, it is exactly that property that also creates some ethical problems. Texts reflect the authors' biases, which get magnified by statistical models. This has unintended consequences for our analysis: If our data is not reflective of the population as a whole, if we do not pay attention to the biases contained, we can easily draw the wrong conclusions, and create disadvantages for our users.

In this talk, I will discuss several types of biases that affect NLP models, their sources, and potential counter measures: (1) Bias stemming from data, i.e., selection bias (if our texts do not adequately reflect the population we want to study), label bias (if the labels we use are skewed) and semantic bias (the latent stereotypes encoded in embeddings); (2) Biases deriving from the models themselves, i.e., their tendency to amplify any imbalances that are present in the data; (3) Design bias, i.e., the biases arising from our (the researchers) decisions which topics to analyze, which data sets to use, and what to do with them. For each bias, I will provide examples and discuss the possible ramifications for a wide range of applications, and various ways to address and counteract these biases, ranging from simple labeling considerations to new types of models.

BIO: Dirk Hovey is an associate professor of Computer Science in the department of marketing at Bocconi University. He received his PhD from the University of Southern California in Los Angeles, where he worked as a research assistant at the Information Sciences Institute. 

He works in Natural Language Processing (NLP), a subfield of artificial intelligence. His research focuses on computational social science. His interests include integrating sociolinguistic knowledge into NLP models, using large-scale statistics to model the interaction between people's socio-demographic profile and their language use, and ethics for data science and algorithmic fairness.

George Em Karniadakis received his SM and PhD from Massachusetts Institute of Technology. He was appointed lecturer in the Department of Mechanical Engineering at MIT in 1987 and subsequently he joined the Center for Turbulence Research at Stanford/Nasa Ames. He joined Princeton University as assistant professor in the Department of Mechanical and Aerospace Engineering and as associate faculty in the program of applied and computational mathematics. He was a visiting professor at Caltech in 1993 in the Aeronautics Department and joined Brown University as associate professor of applied mathematics in the Center for Fluid Mechanics in 1994. After becoming a full professor in 1996, he continues to be a visiting professor and senior lecturer of Ocean/Mechanical Engineering at MIT. He is an AAAS fellow (2018), fellow of the Society for Industrial and Applied Mathematics (2010), fellow of the American Physical Society (2004), fellow of the American Society of Mechanical Engineers (2003) and associate fellow of the American Institute of Aeronautics and Astronautics (2006). He received the Alexander von Humboldt award in 2017, the Ralf E Kleinman award (2015), the J. Tinsley Oden Medal (2013), and the CFD award (2007) from the US Association in Computational Mechanics. His h-index is 103, and he has been cited over 52,000 times.


Abstract:
Karniadakis will present a new approach to develop a data-driven, learning-based framework for predicting outcomes of physical and biological systems, governed by PDEs, and for discovering hidden physics from noisy data. He will introduce a deep learning approach based on neural networks (NNs) and generative adversarial networks (GANs). He will also introduce new NNs that learn functionals and nonlinear operators from functions and corresponding responses for system identification. Unlike other approaches that rely on big data, here we learn from small data by exploiting the information provided by the physical conservation laws, which are used to obtain informative priors or regularize the neural networks. He will demonstrate the power of PINNs for several inverse problems in fluid mechanics, solid mechanics and biomedicine including wake flows, shock tube problems, material characterization, brain aneurysms, etc., where traditional methods fail due to lack of boundary and initial conditions or material properties. He will also present a new NN, DeepM&Mnet, which uses DeepOnets as building blocks for multiphysics problems, and he will demonstrate its unique capability in a 7-field hypersonics application.  

To register and for more information, click here.  

Abstract: Materials used in extreme environments, such as high temperatures, irradiation, and stress, often fail due to rapid defect generation and microstructural evolution, and traditional approaches cannot explore the vast design space needed for next-generation alloys. I will present a machine learning framework powered by massive computing that links individual atomic motion to microstructural evolution. Neural network kinetics models trained on first-principles data map vacancy barrier spectra and capture correlated diffusion in multicomponent alloys, revealing design strategies to suppress radiation damage. At larger scales, simulations uncover dislocation patterning and distinguish between confined and extended slip bands, offering new insight into collective dislocation motion and deformation instabilities. By integrating AI-driven modeling, large-scale computing, and experimental validation, my research goal is to accelerate the discovery of damage-tolerant materials and advance fundamental understanding of defect physics in extreme environments.

Speaker Bio: Penghui Cao is an Associate Professor in Mechanical and Aerospace Engineering at the University of California, Irvine, with a joint appointment in Materials Science and Engineering. He received his PhD in mechanical engineering from Boston University and subsequently worked as a Postdoctoral Associate in the Department of Nuclear Science and Engineering at the Massachusetts Institute of Technology from 2014 to 2018. Dr. Cao's research focuses on understanding the fundamental mechanisms that govern radiation responses and microstructure evolution in materials, and on developing advanced alloys for high-performance nuclear energy systems. His lab advances computational and modeling algorithms, integrates advanced manufacturing techniques to tailor microstructures, and leverages state-of-the-art electron microscopy to characterize and assess underlying mechanisms. He is the recipient of the DOE Early Career Research Program Award and the UCI Samueli School's Mid-Career Award for Faculty Excellence in Research.

Location: Institute for Advanced Computational Science, Seminar Room

*This seminar will be held in-person and online. Zoom link below*

Join Zoom Meeting: https://stonybrook.zoom.us/j/96410717491?pwd=3WGMwbLYNMSbI2IF160VXkvv2JmCQ1.1

Meeting ID: 964 1071 7491
Passcode: 399333