Analyzing the Nonlocality of Sparse Autoencoder Features
Join the next AI-MI Seminar Series on 05/07 at 3:00 PM ET with Xiaoliang Qi, Professor of Physics at Stanford University. This talk explores how sparse autoencoders can be used to extract interpretable features from large language models. Drawing on ideas from holographic duality, the speaker introduces an entropy-based measure to quantify how nonlocal these features are in relation to input tokens—offering new insight into the information dynamics of LLMs. Watch live on YouTube: www.youtube.com/@AIMaterialsInstitute
