"WeightWatcher: Data-Free Diagnostics for Deep Learning"

Abstract: This talk introduces WeightWatcher, a tool for understanding how AI and machine learning models learn. Rather than treating models like deep neural networks or methods such as XGBoost as black boxes, WeightWatcher analyzes patterns in their internal weight and correlation matrices, using ideas from physics and statistics. Well-trained models exhibit clear mathematical signatures associated with good generalization, while poorly trained models display warning signs of overfitting or instability. WeightWatcher provides a simple, science-based framework for evaluating the quality and reliability of modern AI systems. 

Bio: Dr. Martin earned his PhD in theoretical chemistry at the University of Chicago and held an NSF Postdoctoral Fellowship at UIUC. He works on foundational AI research and developed the theory of heavy-tailed self-regularization (HTSR). He created the open-source WeightWatcher project, applying HTSR to assess neural network layer quality. With over 20 years of experience in data science, software engineering, and machine learning, he has contributed to projects at organizations that include Roche, France Telecom, GoDaddy, Aardvark, eBay, eHow, Walmart, GLG, Barclays/BGI, and BlackRock. 

Refreshments and mingling start at 3:20pm

Questions? Contact professor Marcelo Camperi, camperi@usfca.edu