Journal Club, episode 6

Kyle: European Commision White Paper

Lan talks about her idea for a blog post. Continueing her experimentation with fine-tuing GPT2 with Seinfeld scripts, she adapts the work from the Dark Secret of BERT paper to visualize attention weights in the fine-tuned GPT2.

George presents A Unified Approach to Interpreting Model Predictions. A seminal paper both that coins the family of "Additive Feature Attribution Methods", unifying many existing techniques, and puts forth its own, "SHAP". This game theoretic approach combines previous interpretability works with an estimation of Shapley Values.