AI3 Seminar

Meir Feder

Professor, School of Electrical Engineering
Jokel Chair in Information Theory, School of Electrical and Computer Engineering
Tel-Aviv University

Information-Theoretic Framework for Understanding Modern Machine-Learning

Abstract:

Information Theory views learning as universal prediction under log loss, characterized through regret bounds. Unlike the classical results that considered ``small'' model classes and provided uniform regret, the proposed framework provides non-uniform, model dependent bounds utilizing an effective notion of architecture-based model complexity. This complexity is defined by the probability mass or volume of the set of all models in the vicinity of the target model \theta_0, in an informational distance. This volume might be hard to evaluate, yet by local analysis it is related to spectral properties of the expected Hessian or the Fisher Information Matrix at \theta_0, leading to tractable approximations. We argue that successful architectures possess a broad complexity range, enabling learning in highly over-parameterized model classes. The framework sheds light on the role of inductive biases, the effectiveness of stochastic gradient descent (SGD) algorithm, and phenomena such as flat minima. It unifies online, batch, supervised, and generative settings, and applies across the stochastic-realizable and agnostic regimes. Moreover, it provides insights into the success of modern machine-learning architectures, such as deep neural networks and transformers, suggesting that their broad complexity range naturally arises from their layered structure. These insights open the door to the design of alternative architectures with potentially comparable or even superior performance.

Biography:

Meir Feder received the Sc.D. degree in Electrical Engineering and Ocean Engineering in 1987 from the Massachusetts Institute of Technology (MIT) and the Woods Hole Oceanographic Institution (WHOI). After being a Research Associate and a Lecturer at MIT, he joined the School of Electrical Engineering, Tel-Aviv University in 1990, where he is the Jokel Chaired Professor and the former founding head of Tel-Aviv university center for Artificial intelligence and Data science (TAD). Parallel to his academic career, he is closely involved with the high-tech industry: he founded 5 companies, among them Peach Networks (Acq: MSFT) and Amimon (Acq:LON.VTC). Recently, with his renewed interest in machine learning and AI, he co-founded Run:ai (Acq:NVDA), a virtualization, orchestration, and acceleration platform for AI infrastructure, acquired by Nvidia to support its GPU cloud operation.

Prof. Feder received several academic and professional awards including the IEEE Information Theory Society best paper award, the Padovani lectureship, the creative thinking award of the Israeli Defense Forces, and the Research Prize of the Israeli Electronic Industry, awarded by the President of Israel. For the technology he developed in Amimon, he received the 2020 Scientific and Engineering Award of the Academy of Motion Picture Arts and Sciences (OSCAR) and was announced the principal inventor of the technology that attained the 73rd Engineering Emmy Award of the Television Academy.

Location: NCS120


Abstract:

Many real world complex problems are multi-step reasoning tasks. These range from analytic tasks such as answering questions to automation tasks where agents complete tasks on behalf of users.. Evaluation, datasets, and models for such tasks can be unreliable for multiple reasons. (i) Datasets often have annotation artifacts and biases, allowing models to take reasoning shortcuts. Such shortcuts can allow models to make effective guesses -- or, in a sense, cheat -- to achieve high performance without any multi-step reasoning. This issue is further exacerbated for complex tasks because as the number of the required reasoning steps increases, so do the avenues for bypassing those steps. (ii) Models trained on such dataset/s learn to solve the task by taking reasoning shortcuts instead of proper multi-step reasoning. As a result, these models are not robust (reliable) when evaluated in an out-of-distribution evaluation setting. (iii) Lastly, recent works have shown that language models can solve complex multi-step tasks by producing a step-by-step explanation without any training. However, these methods often hallucinate factually incorrect (i.e., unreliable) explanations when posed with knowledge-intensive tasks.

I address these challenges by carefully characterizing the requirements of robust multi-step reasoning and designing reliable evaluation datasets and training methods that necessitate thorough multi-step reasoning. In DiRe, I first formalize and introduce Disconnected Reasoning, i.e., reasoning that allows models to arrive at the correct answer by bypassing necessary reasoning steps, and use this formalization to measure how much multi-step reasoning a model does on a dataset. In MuSiQue, I built a multi-step reasoning dataset for QA from scratch that avoids cheatability via disconnected reasoning, providing a more reliable evaluation. In TeaBReaC, I developed a synthetically generated multi-step QA pretraining dataset designed to force models to avoid disconnected reasoning and learn reliable multi-step reasoning. In IRCoT, I address the reliability of model-generated multi-step reasoning chains by interleaving models' step-by-step reasoning with a step-by-step retrieval from an external corpus, resulting in more factually correct reasoning. Finally, in AppWorld, I built a multi-step reasoning dataset that requires highly interactive problem-solving in an environment carefully designed to ensure models need thorough reasoning to succeed.
Speaker: Harsh Trivedi

Location: NCS 220 or Zoom

https://stonybrook.zoom.us/j/99096379762?pwd=zYCJZQVxRuZd9BboscO4nlodCwsKBr.1
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. 
https://meetings.cshl.edu/meetings.aspx?meet=naisys&year=20  


November 9 - 12, 2020 Virtual
Abstract Deadline: August 14, 2020


Organizers:

Raia Hadsell, DeepMind, United Kingdom
Blake Richards, Mila, Québec AI Institute, Canada
Anthony Zador, Cold Spring Harbor Laboratory

*********

The current COVID-19 situation is challenging and difficult for all of us - we hope this virtual conference will allow colleagues to share and discuss their latest research, while under travel and stay-at-home restrictions.

Because of the ongoing COVID-19/SARS-CoV-2 outbreak, CSHL and the organizers have now reached the difficult decision to restructure the meeting on From Neuroscience to Artificially Intelligent Systems into a virtual meeting scheduled November 9-12, 2020.  NAISys will begin at 10 am (EDT)  on Monday, November 9 and ending with an afternoon session on Thursday, November 12, 2020. Oral sessions will be confined to later morning and afternoon sessions EST to maximize access by participants from around the world. 

*********

Artificial intelligence (AI) and neural networks have long drawn on neuroscience for inspiration. However, in spite of tremendous recent advances in AI, natural intelligence is still far more adept at interacting with the real world in real-time, adapting to changes, and doing so under significant physical and energetic constraints. The goal of this meeting is to bring together researchers at the intersection of AI and neuroscience, and to identify insights from neuroscience that can help catalyze the development of next-generation artificial systems.

Abstracts are welcomed on all scientific topics related to how principles and insights from neuroscience can lead to better artificial intelligence. Topics of interest include but are not limited to network architectures, computing with spiking networks, learning algorithms, active perception, inductive bias, meta-learning, and online learning. Please note that abstracts should be ONE page (~2900 characters).   




Keynote speakers (pending reconfirmation):Yoshua Bengio, Université de Montréal
Yann Lecun, NYU/Facebook


Invited Speakers (pending reconfirmation):Kwabena Boahen, Stanford University
Dmitri Chklovskii, Simons Foundation
Anne Churchland, Cold Spring Harbor Laboratory
Claudia Clopath, Imperial College London, UK
Jim DiCarlo, MIT
Chelsea Finn, Stanford University
Surya Ganguli, Stanford University
Jeff Hawkins, Numenta
Konrad Kording, University of Pennsylvania
Timothy Lillicrap, DeepMind
Yael Niv, Princeton University
Bruno Olshausen, UC Berkeley
Cristina Savin, New York University
Terry Sejnowski, Salk Institute for Biological Studies
Sebastian Seung, Princeton University
Eero Simoncelli, New York University
Sara A. Solla, Northwestern University
David Sussillo, Google Brain
Andreas Tolias, Baylor College of Medicine


New and revised abstracts should be submitted by the resubmission deadline, Friday, August 14. Individuals originally selected for talks should assume they will still be speaking, but we may select additional talks based on the number of invited and selected speakers who cannot reconfirm.

Abstracts should contain only new and unpublished material and must be submitted electronically by the abstract deadline. Selection of material for oral and poster presentation will be made by the organizers and individual session chairs. Status (talk/poster) of abstracts will be posted on our web site as soon as decisions have been made by the organizers.

We are eager to have as many students and postdocs as possible to attend since they are likely to benefit most from this meeting. We have applied for funds from industry and foundations to partially support graduate students and postdocs. Apply in writing stating need for financial support to Catie Carr at carr@cshl.edu. Preference is given to those submitting abstracts. 

All questions pertaining to registration, fees, abstract submission or any other matters may be directed to Catie Carr at carr@cshl.edu.

We anticipate the following support :

National Science Foundation

Social Media:

The designated hashtag for this meeting is #cshlNeuroAI. Note that you must obtain permission from an individual presenter before live-tweeting or discussing his/her talk, poster, or research results on social media. Click the Policies tab above to see our full Confidentiality & Reporting Policy.


Pricing:

Virtual Academic Package: $285
Virtual Graduate Student Package: $175
Virtual Corporate Package: $425

Lab Group Discounts (not departmental or institutional discounts):

Labs registering 4 people: 20% discount off applicable fees
Labs registering 5 people: 25% discount off applicable fees
Labs registering 6 people: 30% discount off applicable fees

To be eligible for lab group discounts, please submit a list of lab members planning to attend in advance of registration to Catie Carr  to establish appropriate discounted fees. Please include a link to your lab web page for verification purposes. Prior payments will be included in the group discount calculation.

IBRO/International Brain Research Organization are generously supporting the participation of a limited number of individuals from US/Canadian Minority Serving Institutions (check eligibility): $25
Limit: 65 attendees / limit per institution: 5 (contact Catie Carr  to confirm eligibility/availability prior to registering) 

Reduced Pricing for Individuals from US/Canadian Minority Serving Institutions (check eligibility): $50

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
West Campus - SAC- Student Activities Center - Ballrooms A & B 100 Nicolls Road Stony Brook NY 11794 Job Fair.jpg The Career Center invites Alumni Employers and Job Seekers to the IT/Computer Science Job and Internship Fair this spring. Job Seekers: A job fair is an opportunity for you to present yourself professionally in person to a potential employer, while showcasing your communication skills. Get more information Alumni Employers: Held in both the fall and spring semesters, this event is ideal for employers looking to fill internship, co-op, part-time and full-time opportunities in the field of information technology (i.e. Software Engineering, Network Administration, Web Development, etc.). Register here to recruit top SBU talent.