Hila Chefer - Transformer Explainability

August 4th, 2022. Columbia University

Abstract:

Transformers have revolutionized deep learning research across many disciplines, starting from NLP and expanding to vision, speech, and more. In my talk, I will explore several milestones toward interpreting all families of Transformers, including unimodal, bi-modal, and encoder-decoder Transformers. I will present working examples and results that cover some of the most prominent models, including CLIP, ViT, and LXMERT. I will then present our recent explainability-driven fine-tuning technique that significantly improves the robustness of Vision Transformers (ViTs). The loss we employ ensures that the model bases its prediction on the relevant parts of the input rather than supportive cues (e.g., background).

Bio:

Hila is a Ph.D. candidate at Tel Aviv University, advised by Prof. Lior Wolf. Her research focuses on developing reliable XAI algorithms and leveraging them to promote model accuracy and fairness.

[cornevanzyl]

As someone who has studied GradCAM in a lot detail, I found this really interesting. I've had my doubts about attention for explainability, I've seen people try many variations in architectures to improve the faithfulness of attention explanations which has casted doubt in my mind on the whole approach but using the gradients is brilliant idea. Your research has renewed my interest in this topic!