Advanced Machine Learning - Co-Teacher

Lecture on Mechanistic Interpretability

The lecture focused on challenges related to interpreting deep neural networks (DNNs) and more specifically transformer-based large language models (LLMs).

The lecture covered a variety of topics related to interpretabilty of AI systems, including:

• The difference between explainability and interpretability focused research
• Feature visualization techniques
• Attention visualization
• Circuit analysis
• Perspectives on the transformer archiecture (the residual stream view)
• The problems of polysemanticity
  • How to extract monosemantic features
• Representation engineering

If interested I can provide slides and lecture notes from the presentation.