AI Innovation Institute

Join a faculty development program to support instructors across campus with navigating/integrating AI in their courses. We're inviting interested faculty to participate in the grant project called Fostering Writing-to-Learn Skills with Critical AI Literacy: A Faculty Development and Student Support Program (funded through the AI3 Institute).

Time commitment and completion requirements :

• Attend four sessions and a final symposium on the following dates/times:

  • Friday, September 12 from 11am - 12:30pm over Zoom

  • Friday, September 26 from 11am - 12:30pm over Zoom

  • Friday, October 10 from 11am - 12:30pm over Zoom

  • Friday, October 24 from 11am - 12:30pm over Zoom

  • Friday, November 14 from 10am - 1pm in Wang 201 - please note that this is an in person session only

• Engage with online materials in Brightspace prior to each of the sessions (mainly to update a syllabus, assignment, or teaching strategy that you can share and discuss at the workshop)

Contact: Shyam Sharma, Christine Fena, and Rose Tirotta-Esposito with questions.

https://docs.google.com/document/d/1b51tvfK0HSOkCW7cwYq2nyyeeHtvBZYC7_XHv7Av8wQ/edit?tab=t.0

Predicting Subjective Attributes in Visual Data - Zijun Wei

ABSTRACT: Recent progress in deep neural networks has revolutionized many computer vision tasks such as image classification, detection and segmentation. However, in addition to excelling in tasks that predict well-defined objective information, human-centered artificial intelligence systems should also be able to model subjective attributes, as defined by human perceptual behavior, that goes beyond the pure physical content of visual data. Example subjective tasks are the prediction of spatial or temporal regions that are interesting to humans (e.g., attract attention or are visually pleasing) and the recognition of subjective attributes (e.g., visually elicited sentiments). Better models for these tasks will improve the human-computer interaction experience in various applications. This thesis investigates several approaches to address the challenges in predicting those subjective attributes in visual data over a diverse set of tasks. I first present a novel framework for real-time automatic photo composition. The framework consists of a cost-effective data collection workflow, an efficient model training pipeline and a lightweight module to account for personalized preferences. Then I develop a novel and general algorithm to detect interesting segments in sequential data, which can be naturally applied to video summarization tasks. Furthermore, I propose methods that learn to represent sentiments elicited by images, in an unsupervised manner, using linguistic features extracted from large scale Web data. To conclude this thesis, I introduce a human-vision-inspired image classification algorithm that also predicts spatial visual attention even though no attention data was used for training it.  

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

OVERVIEW

This workshop, Expanding Horizons in AI with HPC, aims to explore the dynamic intersection of AI and HPC, focusing on how advanced computing can accelerate AI research and applications. As AI models become more complex and data-intensive, traditional computing systems struggle to meet the demand for scalability, efficiency, and speed. HPC offers a solution by providing the necessary infrastructure for training large-scale models, enhancing AI algorithms, and enabling breakthroughs in fields such as deep learning, natural language processing, and autonomous systems.

Through a combination of expert presentations and panel discussions, participants will gain insights into the latest developments in AI-HPC integration. Attendees will also engage in discussions on the future trends, challenges, and ethical considerations surrounding the use of HPC in AI.

The workshop is designed for AI researchers, data scientists, engineers, and HPC professionals seeking to enhance their understanding of how high-performance computing can drive innovation and expand the potential of AI in solving complex, real-world problems.

The workshop will be held at the Wang Center at Stony Brook University.

https://you.stonybrook.edu/hpcai/

PROGRAM

The program features sessions on HPC Architectures for AI, AI Applications in HPC, LLM's in HPC, and AI in HPC Workflows, and open student presentations. The tentative program and list of confirmed speakers is available at https://you.stonybrook.edu/hpcai/program/.

CALL FOR STUDENT PRESENTATIONS & PARTICIPATION

We are excited to offer students the opportunity to present their work in the area of high-performance scientific computing and artificial intelligence at the workshop. We are calling for students to submit their talk proposals (Name + Title) by April 15 to hpc_ai_workshop@stonybrook.edu. The committee will select the best submission to be presented at the workshop. Accepted speakers will be notified by April 22, 2025.

All students, regardless of whether they are presenting, may reach out to hpc_ai_workshop@stonybrook.edu for financial support to cover travel and lodging costs.

REGISTRATION

Registration is available at https://www.eventbrite.com/e/expanding-horizons-in-ai-with-hpc-tickets-1256469978529?aff=oddtdtcreator until May 2nd. The registration fee covers the workshop participation and the social event in the evening of May 9.

Regular registration: $200
Student registration: $100


IMPORTANT NOTE

The registration fee was meant to cover the room rent, catering, and dinner. Thanks to an RF seed grant, we are able to drop the registration fees for SBU students and staff/faculty. We still ask for an informal registration via email to hpc_ai_workshop@stonybrook.edu until April 27, so we can plan for catering and dinner.
Please get in touch with us if you have already registered as an SBU student/faculty/staff member for the workshop so we can handle any reimbursement.

The program is now online at https://you.stonybrook.edu/hpcai/program/.

Virtual Talk: Metadata Matters: Robust Document Classification via Adaptation Methods for Text-driven Public Health by Xiaolei Huang

Zoom link to follow.

Abstract: Document classifiers have been widely applied in solving health-related issues, such as suicide prevention, flu vaccination surveillance and disease diagnosis. However, document metadata including time, gender, age and location has an enormous impact on robustness of 
document classifiers. Language varies across the metadata bringing both challenges and opportunities to build reliable document classifiers. For example, online written language changes over time, and males and females express opinions differently. This talk describes how to use domain adaptation to integrate temporal and user demographic factors into document classifiers. By adapting knowledge of how language varies across the metadata, models can learn generalized representations of language through the metadata-invariant embeddings. 
This approach will lead to metadata-adapted document classifiers and can also extend to personalize classification models by user embedding. 

Bio: Xiaolei Huang is a 4th-year PhD candidate in Information Science at the University of Colorado, Boulder. He is currently a visiting scholar at the Johns Hopkins University. His research interests are in Natural Language Processing, Machine Learning and Public Health. Particularly, he focuses on domain adaptation, cross-lingual transfer learning, user modeling and fairness.

Title: Building foundation models for scientific data Seminar

Speaker: Ruben Ohana, Ph.D. and Michael McCabe, Ph.D - Flatiron Institute, New York

Abstract: Foundation models are very large architectures trained on large-scale datasets and can be used to transfer knowledge from a domain to another. Scientific data, particularly numerical simulations of partial differential equations (PDEs), presents unique challenges due to its complexity and the need for domain expertise to assess prediction quality, complicating the building of the first foundation models in this field. In this talk, we will develop our approach of building foundation models for scientific data, highlighting the requirements and expectations for achieving meaningful results. We will also introduce The Well, a comprehensive collection of datasets encompassing multi-scale simulations of fluid dynamics, astrophysics, and biological systems. The Well serves as a foundation for developing models that generalize across diverse physical phenomena, aiming to accelerate scientific discovery through large-scale learning.

Join Zoom Meeting: https://bnl.zoomgov.com/j/1606898802?pwd=GbbPiLGHlEokDskxjeFheMFWfuboxO.1
Meeting ID: 160 689 8802
Passcode: 281575

Abstract: Artificial Intelligence for Science (AI4Sci) has become a transformative approach in modeling and understanding complex physical systems, encompassing different scales such as atomistic systems and continuum systems. In atomistic systems, AI has shown potential in accelerating simulations, optimizing molecular dynamics, and predicting material and molecular properties through data-driven approaches, enhancing computational efficiency while preserving accuracy. For continuum systems, AI provides powerful tools for solving partial differential equations (PDEs) and learning physical patterns from data, capturing intricate dynamics that govern physical and engineering processes. This work explores AI methods--particularly equivariance for neural networks and neural operators--bridging atomistic and continuum representations. We analyze the implications of incorporating symmetries to improve model robustness and learning efficiency, providing a cohesive AI- driven framework for advancing scientific discovery. The findings aim to underscore the role of AI in enhancing accuracy, applicability, scalability, interopretability, and generalization across scales, from molecular simulations to physical modeling, opening pathways for next- generation applications in computational science. Biography: Wenhan Gao is a third-year Ph.D. student in Applied Mathematics at Stony Brook University, where he works under the supervision of Professor Yi Liu. Wenhan's research focuses on equivariant neural networks, graph neural networks, and AI for partial differential equations. Wenhan's work seeks to leverage the power of symmetries to aid AI models, particularly in fields such as computer vision (image and video generation), physical simulation (modeling climate change), and computational chemistry (drug discovery). He has published papers on the aforementioned topics in leading venues like NeurIPS, Transactions on Machine Learning Research (TMLR), and Journal of Computational Physics (JCP). He also has several preprints under review in leading venues like ICLR and CVPR. In addition to his research, Wenhan has served as a reviewer for top-tier conferences, including ICLR, NeurIPS, ICML, and KDD, and as a lecturer for undergraduate and graduate courses at Stony Brook University. Wenhan was awarded the NeurIPS Travel Award and Excellence in Teaching for Fall 2023.

Abstract: Molecular learning has become an emerging field of AI, driving breakthroughs in drug discovery, protein design, and materials design. For high-stakes scientific tasks, however, predictive accuracy alone is not sufficient: models must also be interpretable and trustworthy. Our work aims to study molecular learning under a unified explainability perspective across two major model families: Graph Neural Networks (GNNs) and Large Language Models (LLMs).

GNNs are natural choices for molecular graphs and achieve strong performance on many molecular tasks. To enhance explainability, many GNN explanation methods have been proposed and work well for 2D GNNs. However, 3D GNNs introduce two key challenges: producing chemically meaningful substructures and reducing fidelity loss caused by dense geometric graphs. To address these challenges, I present two methods. 3DGraphX decomposes dense 3D graphs into chemically meaningful 3D motifs, enabling compact explanations that align with chemical intuition. EDMA introduces an energy-based discrete mask approximation approach to reduce the discrepancy between the soft mask optimized during training and the hard mask used for explanation, improving explanation fidelity.

LLMs present different characteristics and challenges compared with GNNs. LLMs can provide a certain level of explanation through step-by-step reasoning, and their natural-language outputs are easy for humans to understand and interpret. However, because LLMs are trained for general-purpose tasks, their performance on scientific tasks often lags behind specialized GNNs. To improve performance, existing methods guide LLMs by providing suggestions through brief feedback, retrieval-augmented generation (RAG), or planner agents. However, these approaches face several limitations, such as vague guidance, introduced bias problems, and high computational cost. To fill the gap, I propose RL-Guider, a lightweight reinforcement-learning agent that converts evaluation feedback into input-specific guidance for molecular optimization. RL-Guider improves over time by accumulating historical experience and transfers efficiently across different LLMs while preserving interpretability.

Together, these efforts aim to provide explanations that are scientifically meaningful and faithful, while also preserving or improving performance on molecular tasks to better meet real scientific needs.

Speaker: Xufeng Liu

Location: New Computer Science-1-Room 115

Learn about a new feature in Gemini that can interact with Google Workspace

Learn how to use Gemini for Google Workspace. We'll move past basic prompts and show you how to use Gemini as an assistant that works with the other parts of Google Workspace from Gmail, Docs, and Calendar.

Register here.