TITLE: Sampling Using Langevin Diffusions Beyond the Worst-Case by Andrej Risteski of CMU


ABSTRACT: Many tasks involving generative models involve being able to sample from distributions parametrized as p(x) = e^{-f(x)}/Z where Z is the normalizing constant, for some function f whose values and gradients we can query. This mode of access to f is natural -- for instance sampling from posteriors in latent-variable models. Classical results show that a natural random walk, Langevin diffusion, mixes rapidly when f is convex. Unfortunately, even in simple examples, the applications listed above will entail working with functions f that are nonconvex.

We exhibit instances where Langevin diffusion (combined with other tools) can provably be shown to mix rapidly in instances of relevance in practice: distributions p that are multimodal, as well as distributions p that have a natural manifold structure on their level sets. 
Join Stony Brook University's Center for Excellence in Learning and Teaching (CELT) for a boot camp on how to use AI to enhance your teaching and courses. This event will demonstrate how ChatGPT, Microsoft Copilot, NotebookLM, and other generative AI platforms can support you in crafting learning objectives, writing exam questions, composing rubrics, and designing course content such as lesson plans, in-class activities, instructional videos, and more.

https://stonybrook.zoom.us/j/92511854285?pwd=QRTHfULqHMWxJYoVyt3piOhNxWLfvs.1
Abstract: Datalog is a powerful language for expressing recursive computations through rules: Horn clauses in first order logic. Although effective at expressing queries over existential properties, Datalog and many of its popular implementations struggle with queries that involve more complex aggregates, requiring users to apply verbose, non-composable, and/or inefficient workarounds. Recent work on lattice-based datalogs addresses many of these concerns for aggregates that can be encoded as lattices (e.g., min or max), but more general aggregates like count remain problematic. In this talk, I will argue that this is not a fundamental limitation of Datalog, but rather from its model of truth: Both datalog semantics and evaluation rules make heavy use of the fact that insertion is both monotone and idempotent. Once a fact is known to be true, it can not be retracted, nor can further discoveries of the same fact alter its truth. Monotonicity is critical for forward progress under Datalog's ``open world'' model, as it allows us to safely assert the truth of a body. Meanwhile, idempotence makes it easier to reason about evaluation, as we need only guarantee that each head atom will be derived at-least-once. Unfortunately, more general aggregates like sum() are neither idempotent, nor monotone. I will introduce Hedgelog, a strict generalization of Datalog that uses general monoids as a basis for truth. I will show that this generalization remains compatible with Datalog's open world model, how it enables cleaner and more composable datalog programs, and how the underlying monoid relations open the door to interesting datastructure-level optimizations.

Bio: Oliver Kennedy is an associate professor at the University at Buffalo. He earned his PhD from Cornell University in 2011 and now leads the Online Data Interactions (ODIn) lab, which operates at the intersection of databases and programming languages. Oliver is the recipient of an NSF CAREER award, an IEEE Region 1 Technological Innovation Award, UB's Exceptional Scholar Award, and several UB SEAS teaching awards. Oliver is also one of the founding board members of Breadcrumb Analytics. Several of Oliver's papers have been invited to Best of compilations from SIGMOD and VLDB. The ODIn lab is currently exploring (i) how we can leverage database techniques like incremental view maintenance to make compilers faster, (ii) how to make it easier for data scientists to track how sources of uncertainty, ambiguity, and/or bias affect analyses, and (iii) how to streamline the interfaces --- both human and software --- between different tools for data science, like python, sql, and spreadsheets.

Location: NCS 120

The event will take place on Zoom and will feature two distinguished guest speakers: SBU alumnus, Velchamy Sankarlingam, president of Product and Engineering at Zoom, and Simeon Ananou, vice president for Information Technology and CIO at Stony Brook University. The discussion will be moderated by Haresh Gurnani, dean of the College of Business at Stony Brook University.

Exploring AI's Impact on Communication and Connection

Artificial Intelligence (AI) has rapidly evolved, becoming an integral part of various industries, including education and business. This event aims to delve into how AI is reshaping the way we learn and work, particularly in enhancing communication and fostering human connections. Velchamy Sankarlingam, an SBU alumnus and a key figure at Zoom, will share his insights on how AI-driven tools are revolutionizing virtual communication platforms, making interactions more seamless and effective.

Simeon Ananou, with his extensive experience in information technology, will provide a perspective on how AI is being integrated into educational institutions to improve learning outcomes and administrative efficiency. His role at Stony Brook University places him at the forefront of implementing innovative technologies that benefit both students and staff.

A Conversation Led by Expertise

Dean Haresh Gurnani, known for his leadership and expertise in business education, will guide the conversation, ensuring that the discussion remains focused on the practical implications of AI. He will explore how AI is not only boosting productivity but also enriching overall experiences in the workplace and educational settings. The event will include an interactive Q&A session, allowing attendees to engage directly with the speakers and gain deeper insights into the topics discussed.

As AI continues to develop, events like this are crucial for understanding its impact and potential. Stony Brook University's College of Business is committed to providing platforms for such important discussions, fostering an environment where innovation and education intersect.

This event is open to all. Please visit https://www.givecampus.com/schools/StonyBrookUniversity/events/artificial-intelligence-reshaping-learning-and-work to register.

Abstract: Self-supervised representation learning (SRL) has emerged as a pivotal advancement in machine learning, offering high-quality data representations without the need for labeled datasets. While SRL has demonstrated enhanced adversarial robustness compared to supervised learning, its resilience against other attack types, particularly backdoor attacks, remains an open question. Recent studies have revealed potential vulnerabilities in SRL, underscoring the necessity for a comprehensive security analysis. However, existing research often extrapolates attacks from supervised learning paradigms, neglecting the unique challenges and opportunities inherent to self-supervised mechanisms.

This thesis proposal aims to address three critical objectives in the realm of self-supervised learning: (1) exploring novel attack vectors, (2) implementing and evaluating practical attacks, and (3) developing robust countermeasures. We focus on two key SRL paradigms: Contrastive Learning and Diffusion Models. For Contrastive Learning, we synthesize existing security vulnerabilities and introduce innovative attack vectors, such as CTRL, to uncover distinctive risks. We conduct a comparative analysis of contrastive and supervised learning approaches in their defense against these threats, exploring potential safeguards and highlighting the limitations of current protective measures in self-supervised contexts. Regarding Diffusion Models, we demonstrate inherent vulnerabilities in their application to adversarial purification.

Our research aims to illuminate the unique challenges posed by emerging attack vectors in self-supervised learning, fostering technical advancements to address underlying security risks in real-world applications. By contributing to the development of more resilient and secure self-supervised representation learning systems, we seek to enhance their reliability and trustworthiness in practical scenarios. This comprehensive examination of SRL's security landscape will provide valuable insights for the broader machine-learning community and pave the way for more robust AI systems.

Join here.
Speaker Petar Djuric Refreshments will be provided Deep Gaussian processes: Theory and applications Petar M. Djurić Department of Electrical and Computer Engineering Stony Brook University Abstract: Gaussian processes are an infinite-dimensional generalization of multivariate normal distributions. They provide a principled approach to learning with kernel machines and they have found wide applications in many fields. More recently, with the advance of deep learning, the concept of deep Gaussian processes has emerged. Deep Gaussian processes can be viewed as multilayer hierarchical organizations of Gaussian processes that are equivalent to infinitely wide multiple layer neural networks. Deep Gaussian processes have improved capacity for prediction and classification over standard Gaussian processes, while models based on them continue to allow for full Bayesian treatment and for applications when the amount of available data is limited. The theory of recent progress in deep Gaussian processes will be presented and some applications will be provided. Biosketch: Petar M. Djurić received the B.S. and M.S. degrees in electrical engineering from the University of Belgrade, Belgrade, Yugoslavia, respectively, and the Ph.D. degree in electrical engineering from the University of Rhode Island, Kingston, RI, USA. He is a SUNY Distinguished Professor and currently, he is a Chair of the Department of Electrical and Computer Engineering, Stony Brook University, Stony Brook, NY, USA. Djurić was a recipient of the IEEE Signal Processing Magazine Best Paper Award in 2007 and the EURASIP Technical Achievement Award in 2012. From 2008 to 2009, he was a Distinguished Lecturer of the IEEE Signal Processing Society. He was the Editor-in-Chief of the IEEE Transactions on Signal and Information Processing over Networks (2015-2018). Djurić is a Fellow of IEEE and EURASIP

Abstract:

It is known that models like large language models (LLMs) can often suggest colloquial plans given verbal descriptions of tasks, yet they are unable to reliably provide executable and verifiable plans given formally specified environments. In this talk, I will discuss a strand of efforts to have LLMs generate accurate and explainable plans in textual simulations. Instead of directly generating the plan or actions, LLMs are prompted to generate Planning Domain Definition Language (PDDL) that specifies the environment (domain file) and the task (problem file), which can then be deterministically solved with an off-the-shelf planner. In a 3-phase study, my collaborators and I first observed that it is possible but very challenging for LLMs to generate long-form code such as PDDL domain and problem files given textual specifications. Next, we devise methodologies for LLMs to iteratively generate and refine problem files while exploring a partially-observed, simulated, textual environment. Finally, we show that domain files are even more difficult to generate correctly, even on well-established planning tasks such as BlocksWorld. Finally, I will discuss ongoing efforts to improve said ability of structured generation and promising frontiers to explore.

Bio:
Li Harry Zhang is an assistant professor at Drexel University, focusing on Natural Language Processing (NLP) and artificial intelligence (AI). He obtained his PhD degree from the University of Pennsylvania advised by Prof. Chris Callison-Burch. Prior, he obtained his Bachelor's degree at the University of Michigan mentored by Prof. Rada Mihalcea and Prof. Dragomir Radev. His current research uses large language models (LLMs) to reason and plan via symbolic and structured representations. He has published more than 20 peer-reviewed papers in NLP and AI conferences, such as ACL, EMNLP, and AACL, that have been cited more than 1,000 times. He also consistently serves as Area Chair, Session Chair, and reviewer in those venues. Being a musician, producer, and content creator having over 50,000 subscribers, he is also passionate in the research of AI music and creativity.

The AI Community at Stony Brook University is proud to announce Datathon 2026.

Dive into data analysis and AI/ML, and get ready to build something big. In this year's underwater-themed event, enjoy a weekend of data analysis, hacking, networking, fun activities, and minigames.

Whether you're a seasoned developer, data scientist, designer, or completely new to hacking, this event is your chance to collaborate, learn data science, and create something impactful with data and AI/ML.

What is Datathon?

AI Community's Datathon is the premier data science competition at Stony Brook University, bringing together students of all skill levels for a weekend of data exploration, analysis, and innovation. Just like a typical hackathon, you will be using your skills to build your dream project.

Unlike a regular hackathon, Datathon is focused on data science. You will be given a set of data to work with, analyze, and apply to your project. You can also find your own data to use. Your project will be presented to a panel of judges consisting of professors and industry professionals!

Who Can Participate

  • Students of all skill levels and majors are welcome.
  • Come with a team or find one at the event or on Discord.
  • This event is open to SBU and non-SBU students.

* Non-SBU Undergraduate Students are ineligible to receive prizes
* You must be 18+ or older (Excludes minors who are active SBU students)

Location: SAC Ballroom B

Register here.