AI Innovation Institute

University Libraries Presents:
Join librarian Christine Fena for an interactive workshop that invites you to explore AI tools first hand, not just as users, but as critical investigators.
Through playful experimentation and collaborative discovery, you'll uncover inherent biases, probe algorithmic flaws, and gain a deeper understanding of AI's limitations and societal impacts.

RSVP on SBEngaged

Location: Melville Library, Central Reading Room, Lab B

Abstract: The capacity to adapt machine learning models to various contexts, information, and objectives is particularly valuable. In this thesis, I focus on developing Class Conditional Guided Models. These are models that can be adaptively biased towards a class of interest via a conditional input. My primary focus lies in the efficiency of these models. They are constructed to require training only once, with the ability to quickly and conveniently adapt during testing time without necessitating fine-tuning or retraining.
Firstly, I propose RelationVAE, a novel generative model designed for few-shot scenarios, utilizing the prior knowledge of class similarity relationships. RelationVAE is designed to condition on the embeddings of the neighbor classes (i.e. classes with similarity relationships), to generate more reliable samples by making them more similar to the neighbor class. This enables adaptation of the generative model to the provided prior knowledge about class relationships.
As a second focus, I introduce scGAN, a shadow segmentation technique that enables adaptation to varying shadow distributions in different testing environments. scGAN is designed to condition on a sensitivity parameter, a scalar, to control the amount of the shadow detected. In the testing phase, the parameter is set to appropriate values, allowing the model to quickly adapt to specific test environments.
In my third contribution, I propose S-SEG, a methodology for fine-grained counting allowing adaptation to different granularities of fine-grained classes. In fine-grained problems, the distinction between classes is subtle and inconsistent across images, leading to variations in the granularity of the target class from one image to another. S-SEG is designed to be conditioned on an additional input, the sensitivity parameter, to control the granularities of the target class during inference.
My fourth contribution is a text-to-image synthesis method which allows controlling the number of the generated objects of a target class. I propose to generate an intermediate condition, the density map, which reflects the number of objects, together with their layout. This intermediate condition is used to effectively guide the generative model to generate objects with accurate counts.

Speaker: Vu Nguyen

Zoom: https://stonybrook.zoom.us/j/97114455337?pwd=Z4rB9dWcstlahUIs8PRrvQ9b2ZK2Df.1
Meeting ID: 971 1445 5337
Passcode: 272300

Abstract: Machine learning (ML) systems fueled by neural networks have entered our daily lives and led to scientific breakthroughs, but many open questions remain. After a nod toward the question of rigor with ML and recent progress, I'll turn to the theory of neural networks. I will argue that understanding neural networks inevitably leads to ideas from field theory (FT), which was already realized in the simplest case in the 1990s, and I will review some essential FT-for-NN results. I will then propose that the connection might be more general, an NN-FT correspondence of sorts, with neural networks providing a way to define a field theory. I'll end with comments on known results including the origin of interactions and various symmetries, but I will also list some open questions. The apparent non-sequitur in the title will be used as a rhetorical device to explore where we are and where we'd like to go.

https://scgp.stonybrook.edu/calendar/full-calendar

Abstract: Artificial intelligence (AI) is rapidly transforming scientific discovery, enabling breakthroughs in areas ranging from drug discovery to modeling complex physical systems. In the life sciences, AI has traditionally been applied to prediction tasks such as classifying molecules as toxic or non-toxic, estimating drug properties, or solving partial differential equations. These discriminative models have proven powerful, but they are inherently limited to mapping existing inputs to deterministic outputs. A new wave of methods is shifting the paradigm from discrimination to generation: creating new possibilities, such as generating novel molecules or designing new drugs. By reframing AI as both a predictive and generative engine, this shift offers new pathways for accelerating discovery and innovation in life sciences at an unprecedented scale. This talk will cover several aspects of AI for Science (AI4Sci), beginning with advances in discriminative models for molecular systems and solving PDEs, and then turning to generative approaches, including diffusion models for 3D molecular generation and large language models for drug editing. Together, these developments illustrate how moving from prediction to creation is redefining what AI can contribute to science.

Bio: Wenhan Gao is a fourth-year Ph.D. student in Applied Mathematics under the supervision of Professor Yi Liu. He was also a Staff Research Scientist Intern at VISA Research, where he worked on large language models (LLMs) and multi-agent systems for commerce. Wenhan's research focuses on AI for Science (AI4Sci), with a particular emphasis on generative AI. His work looks deep into the fundamental mechanisms of AI models when applied to scientific tasks, and he strives to incorporate established scientific priors, such as symmetry, into model design. He has published papers as a first or corresponding author in leading AI and computational venues, including ICLR, ICML, NeurIPS, TMLR, ACL, and the Journal of Computational Physics. In addition to his research, Wenhan has served as a reviewer and oral session chair for top AI conferences and as a lecturer for both undergraduate and graduate courses at Stony Brook University.

Location: IACS Seminar Room or Zoom

This seminar will take place in person and online*

Join Zoom Meeting: https://stonybrook.zoom.us/j/91670093552?pwd=2EcniXqPZLTpa4ZBKRs1zAjYqs1LS0.1

Meeting ID: 916 7009 3552
Passcode: 434045

Abstract: Modern technologies enable enhanced integrity and privacy guarantees not just for data, but also for computation. This is perhaps most emphatically demonstrated by the steady rise of zero-knowledge proofs, which are short certificates that attest to the correctness of computations (e.g., an age verification check) without revealing any secret inputs (e.g., the birth date on a digital ID). This subtly powerful technology enables anonymous credentials, privacy-preserving machine learning, anonymous blockchains, and much more--making the question of efficient zero-knowledge proofs fundamental to modern secure systems. Echoing Moore's law for computing, zero-knowledge proofs have improved on this front by ten orders of magnitude in the last two decades. In this talk, I will discuss our work on overcoming a key bottleneck that has emerged in this development: memory efficiency.

Speaker: Abhiram Kothapalli is a postdoctoral scholar at the University of California, Berkeley, hosted by Sanjam Garg. He is a recent graduate of Carnegie Mellon University, where he earned his Ph.D. in Computer Science, advised by Bryan Parno. Previously, he was at the University of Illinois at Urbana-Champaign, where he earned his B.S. in Computer Science and B.S. in Mathematics. Kothapalli's research develops cryptographic techniques aimed at scaling expressive privacy and integrity guarantees across the internet.

Location: NCS 120

Abstract: Sea ice is crucial to Earth's climate, Arctic communities, and ecosystems, yet climate change is driving significant losses, threatening polar stability. Quantifying the long-term impacts of a declining sea ice cover requires tools which improve climate-timescale prediction and bring new understanding of climate interactions. In this talk, I discuss how meeting this challenge requires a multi-disciplinary approach. Climate models, while essential, suffer from systematic biases due to missing or inaccurate physics, leading to uncertainty in future projections. I show how data assimilation (DA) offers a statistical framework for integrating satellite observations with climate models to quantify systematic sea ice model errors. Using convolutional neural networks (CNNs), we can learn these errors based on the model's atmospheric, oceanic, and sea ice conditions--what I term a state-dependent representation of the error. This approach enables real-time corrections to subsequent model simulations, which systematically reduces global sea ice biases. I highlight key successes and challenges in developing this hybrid ML+climate modeling framework, including transfer learning to enhance online generalization of ML models, and new methods for integrating Python-based ML frameworks with Fortran climate model code. Finally, I introduce GPSat, a scalable Gaussian process-based tool for reconstructing complete sea ice fields from sparse satellite altimetry data. Together, the DA+ML framework and GPSat offer future opportunities for improving targeted model physics errors for more robust climate simulation.


IACS Seminar Speaker: William Gregory, Princeton University

Location: IACS Seminar Room

Abstract: Recent studies have highlighted the vulnerability of Natural Language Processing (NLP) and Vision-Language Models (VLMs) to backdoor attacks, posing significant security risks. Understanding these attack strategies is crucial for assessing model robustness and developing effective defenses. This thesis proposal aims to investigate the vulnerability of language and vision-language models, analyze abnormal behaviors in backdoor-attacked models, and develop defense methods to enhance safety of modern machine learning models at deployment.


We investigate the internal mechanisms of backdoored NLP models, identifying a distinct attention focus drifting phenomenon, where trigger tokens hijack attention regardless of the input context. Through comprehensive qualitative and quantitative analysis, we provide insights into the underlying mechanisms that enable backdoor attacks. Building on these insights, we propose detection methods to differentiate backdoored models from clean ones, through inspecting both the attention distribution and the model predictions. To better understand the vulnerability, we develop advanced backdoor attack strategies targeting language models in classification tasks. For BERT variants, we introduce Trojan Attention Loss (TAL), a novel method that directly manipulates attention patterns to enhance backdoor effectiveness, ensuring stealth and robustness. Vision-Language Models have demonstrated strong performance in recent years. Yet their vulnerability is largely underexplored. We investigate advanced backdoor attack strategies on Vision-Language Models, focusing on image-to-text generation tasks. We demonstrate how backdoors can be embedded in complex multimodal tasks while maintaining semantic integrity under poisoned inputs. Additionally, we propose innovative techniques for injecting backdoors without requiring access to the original training data, expanding the feasibility of real-world attacks.

This proposal provides novel insights into the internal mechanisms of backdoored models, propose effective detection strategies, and develop advanced attack techniques that expose critical vulnerabilities. These findings underscore the urgent need for robust security measures to defend against emerging backdoor threats in deep learning models. The results have been published in top venues including ICLR, ECCV, NAACL, EMNLP, etc.

Speaker: Weimin Lyu


Zoom link: https://stonybrook.zoom.us/j/99880605139?pwd=cfWbRG6n9v3GXEa7OqvXa5cOp5eLBv.1
Meeting ID: 998 8060 5139
Passcode: 843302