Stanford Seminar 2022 - Transformer Circuits, Induction Heads, In-Context Learning
Stanford University via YouTube
Overview
Explore the fascinating world of mechanistic interpretability in neural networks through this Stanford seminar. Delve into the concept of neural network parameters as compiled computer programs and learn how to reverse engineer them into human-understandable algorithms. Focus on transformer language models and discover the significance of "induction head circuits" in enabling in-context learning. Examine how these circuits allow models to repeat text, translate, and mimic functions from earlier context. Understand the pivotal role of induction heads in driving sharp phase changes during the learning process, impacting loss curves and model learning trajectories. Gain insights from Chris Olah, co-founder of Anthropic and leader of their interpretability efforts, as he shares his expertise on AI safety and large model interpretation.
Syllabus
CS25 I Stanford Seminar 2022 - Transformer Circuits, Induction Heads, In-Context Learning
Taught by
Stanford Online