Inside-Out Interpretability: Training Dynamics in Multi-Layer Transformer

Explore a comprehensive lecture from UC Berkeley's EECS department featuring Meta AI Research Scientist and Senior Manager Yuandong Tian as he delves into the intricate world of transformer model interpretability. During this 67-minute presentation, gain deep insights into the training dynamics of multi-layer transformers, examining their internal mechanisms and behavioral patterns. Learn about cutting-edge approaches to understanding how these complex neural networks develop and function from the inside out, with particular focus on their training evolution and architectural implications. Perfect for those interested in advanced machine learning concepts, neural network architecture, and the technical foundations of modern AI systems.