Optimization and Machine Learning - presented by Yifan Sun

Abstract: Optimization is a growing topic of interest in the machine learning community. It starts out as an option to check in Tensorflow (SGD? Adam? Adagrad?), but as we get more into the how and why of these options, we uncover many fundamental principles relating to operations research, control theory, and dynamical systems, dating back as far as the Cold World era. 

In this talk I will give a broad overview of some of the important optimization themes in machine learning. I will try to give connections between tools we are used to seeing in popular packages 
and fundamental optimization concepts like duality, convexity, contractive operators, etc. While we cannot hope to completely cover this diverse research area, I hope to provide a glimpse of this exciting research area that is permeating more and more into the machine learning world. 

Bio: Yifan Sun received her PhD in Electrical Engineering from the University of California Los Angeles in 2015, with research focusing on convex optimization and semidefinite programming. She was then Technicolor Research and Innovation, focusing on machine learning and 
data science applications. More recently, she completed two postdocs focusing on optimization, at the University of British Columbia in Vancouver, Canada and INRIA, in Paris, France.

University Libraries Presents: Do you want more control over when to use AI features in your Google apps and when browsing online? Are you tired of Gmail trying to finish your sentences for you? Drop in to this brief tutorial, which will walk through the settings you can turn on and off in various Google products to keep your agency where you want it.

Register here for the Zoom session.

A talk by Jerome Zhengrong Liang entitled, Machine Learning from Original Images to Texture Patterns: A Paradigm Shift from Non-Medical Application to Medical Diagnosis. Abstract: Artificial intelligence (AI) research for medical diagnosis started soon after human began to use computer, initially called artificial neural network (ANN) and now convolutional neural network (CNN). ANN has been mainly explored to classify the experts' handcrafted features from the original (or raw) images, while CNN has been mainly explored directly on the raw images for both tasks of extracting abstract features and classifying the features. Experimental evidences have been shown that CNN can be trained by a large number of the raw images with experts' scores (or labels) to match or even surpass the experts' performance for both non-medical and medical diagnosis applications. However, the performances of the CNN models as well as the experts on medical diagnosis dropped dramatically when the labels of the raw images were replaced by the corresponding medical pathological reports. Accumulated medical knowledge reveals that the lesion heterogeneity is a footprint of lesion evolution and ecology, and the heterogeneity is an indicator of lesion progress and response to medical intervention. The heterogeneity can be reflected by the image contrast distribution (or texture patterns) across the lesion volume. Image textures have been shown as an effective descriptor of the lesion heterogeneity for computer-aided diagnosis. Can we map the raw images into texture patterns (or images) and train CNN to learn from the texture images? This question is the central theme of this presentation with application to CT Colonography or virtual colonoscopy, a game from AlphaGo to PolypGo. Bio: Jerome Zhengrong Liang, PhD, IEEE Fellow Imaging Research and Informatics Laboratory Department of Radiology, Stony Brook University
The Natural Language Processing Reading Group at Stony Brook University meets weekly to discuss recent research papers in NLP and related fields.
Join the Google Group here.

Memorization Dynamics in Knowledge Distillation for Language Models

Abstract: Knowledge Distillation (KD) is increasingly adopted to transfer capabilities from large language models to smaller ones, offering significant improvements in efficiency and utility while often surpassing standard fine-tuning. Beyond performance, KD is also explored as a privacy-preserving mechanism to mitigate the risk of training data leakage. While training data memorization has been extensively studied in standard pre-training and fine-tuning settings, its dynamics in a knowledge distillation setup remain poorly understood. In this work, we study memorization across the KD pipeline using three large language model (LLM) families (Pythia, OLMo-2, Qwen-3) and three datasets (FineWeb, Wikitext, Nemotron-CC-v2). We find: (1) distilled models memorize significantly less training data than standard fine-tuning (reducing memorization by more than 50%); (2) some examples are inherently easier to memorize and account for a large fraction of memorization during distillation (over ~95%); (3) student memorization is predictable prior to distillation using features based on zlib entropy, KL divergence, and perplexity; and (4) while soft and hard distillation have similar overall memorization rates, hard distillation poses a greater risk: it inherits 2.7 times more teacher-specific examples than soft distillation. Overall, we demonstrate that distillation can provide both improved generalization and reduced memorization risks compared to standard fine-tuning.

Location: NCS 220

Zoom Link: https://stonybrook.zoom.us/j/92942833294?pwd=927683
Abstract: To be helpful assistants, AI agents must be aware of their own capabilitiesand limitations. This includes knowing when to answer from parametricknowledge versus using tools, when to trust tool outputs, and when toabstain or hedge. Such capabilities are hard to teach through supervisedfine-tuning because they require constructing examples that reflect theagent's specific capabilities. We therefore propose a radically new approachto teaching agents what they know: collaborative self-play. We constructmulti-agent collaborations in which the group is rewarded for collectivelyarriving at correct answers. The desired meta-knowledge emerges fromthe incentives built into the structure of the interaction. We focus on smallsocieties of agents that have access to heterogeneous tools (corpus-specificretrieval), and therefore must collaborate to maximize their success withminimal effort. Experiments show that group-level rewards for multiagent communities can induce policies that transfer to improve tool use andselective prediction in single-agent scenarios. Speaker: Syed Mostofa Monsur Location: Old Computer Science Building - CS2311
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 PhD program or (2) receive permission from the instructors.

Each seminar will consist of multiple short talks (around 10 minutes) by multiple people. Students can register for 1 credit for CSE 656. Registered students must attend and present a minimum of 2 or 3 talks. Everyone else is welcome to attend. Fill in https://forms.gle/pCVXovgfMfQwGqG38 to subscribe to our mailing list for further announcement.
Title: Cultural Biases, World Languages, and User Privacy in Large Language Models
Abstract: In this talk, I will highlight three key aspects of large language models: (1) cultural bias in LLMs and pre-training data, (2) decoding algorithm for low-resource languages, and (3) human-centered design for real-world applications.

The first part focuses on systematically assessing LLMs' favoritism towards Western culture. We take an entity-centric approach to measure the cultural biases among LLMs (e.g., GPT-4, Aya, and mT5) through natural prompts, story generation, sentiment analysis, and named entity tasks. One interesting finding is that a potential cause of cultural biases in LLMs is the extensive use and upsampling of Wikipedia data during the pre-training of almost all LLMs. The second part will introduce a constrained decoding algorithm that can facilitate the generation of high-quality synthetic training data for fine-grained prediction tasks (e.g., named entity recognition, event extraction). This approach outperforms GPT-4 on many non-English languages, particularly low-resource African languages. Lastly, I will showcase an LLM-powered privacy preservation tool designed to safeguard users against the disclosure of personal information. I will share findings from an HCI user study that involves real Reddit users utilizing our tool, which in turn informs our ongoing efforts to improve the design of AI models.
Bio:

Wei Xu is an Associate Professor in the College of Computing and Machine Learning Center at the Georgia Institute of Technology, where she is the director of the NLP X Lab. Her research interests are in natural language processing and machine learning, with a focus on Generative AI, robustness and fairness of large language models, multilingual LLMs, as well as AI for science, education, accessibility, and privacy research. She is a recipient of the NSF CAREER Award, Google Academic Research Award, CrowdFlower AI for Everyone Award, Best Paper Awards and Honorable Mentions at COLING'18, ACL'23, ACL'24. She also received research funds from DARPA and IARPA. She is currently an executive board member of NAACL. Join Zoom Meeting https://stonybrook.zoom.us/j/98855994362?pwd=F2qnpwL85fhCBHAEW9ZBpXihfwGHsj.1 (ID: 98855994362, passcode: 172797) Join by phone (US) +1 646-876-9923 (passcode: 172797) Joining instructions: https://www.google.com/url?q=https://applications.zoom.us/addon/invitation/detail?meetingUuid%3DuDJcUTvyQueZkCaUSAwFlg%253D%253D%26signature%3Da3d49e0f7f2e74e7130f7308c74bd85ba7b99587b98ba2e34238bb657ca51a09%26v%3D1&sa=D&source=calendar&usg=AOvVaw2jTn5cjfRG8vXU8KHHlU2Y Meeting host: H.Andrew.Schwartz@stonybrook.edu

Join Zoom Meeting:
https://stonybrook.zoom.us/j/98855994362?pwd=F2qnpwL85fhCBHAEW9ZBpXihfwGHsj.1


The International Conference on Learning Representations (ICLR) is the premier gathering of professionals dedicated to the advancement of the branch of artificial intelligence called representation learning, but generally referred to as deep learning.



ICLR is globally renowned for presenting and publishing cutting-edge research on all aspects of deep learning used in the fields of artificial intelligence, statistics and data science, as well as important application areas such as machine vision, computational biology, speech recognition, text understanding, gaming, and robotics.

ICLR is one of the fastest growing artificial intelligence conferences in the world. Participants at ICLR span a wide range of backgrounds, from academic and industrial researchers, to entrepreneurs and engineers, to graduate students and postdocs.

The rapidly developing field of deep learning is concerned with questions surrounding how we can best learn meaningful and useful representations of data. ICLR takes a broad view of the field and includes topics such as feature learning, metric learning, compositional modeling, structured prediction, reinforcement learning, and issues regarding large-scale learning and non-convex optimization.

A non-exhaustive list of relevant topics explored at the conference include:



  • Unsupervised, Semi-supervised, and Supervised Representation Learning
  • Representation Learning for Planning and Reinforcement Learning
  • Metric Learning and Kernel Learning
  • Sparse Coding and Dimensionality Expansion
  • Hierarchical Models

  • Optimization for Representation Learning
  • Learning Representations of Outputs or States
  • Implementation Issues, Parallelization, Software Platforms, Hardware
  • Applications in Vision, Audio, Speech, Natural Language Processing, Robotics, Neuroscience, or Any Other Field


For more information or registration, please visit the official website.