Скачать или смотреть Vision Mechanistic Interpretability - MATS Talk Summer 2024

This video features Sonia Joseph's MATS talk on vision mechanistic interpretability, presented on August 10th. It's the first part of a workshop series, best suited for intermediate-level viewers with some familiarity with mechanistic interpretability concepts. Sonia introduces the Prisma library for multimodal mechanistic interpretability and demonstrates key techniques for analyzing Vision Transformers (ViTs). (Heads up this is a workshop recording so the audio is a bit raw.)

Links to notebooks in video:
Powerpoint of Presentation: https://docs.google.com/presentation/...
Notebook I: ViT Main Demo: https://colab.research.google.com/dri...
Notebook II: Emoji Logit Lens: https://colab.research.google.com/dri...
Notebook III: Interactive Attention Head Visualization: https://colab.research.google.com/dri...

Other links:
Prisma Library Less Wrong Post: https://www.lesswrong.com/posts/kobJy...
Prisma repo: https://github.com/soniajoseph/ViT-Pr...
ARENA curriculum: https://github.com/callummcdougall/AR...

Timestamps:
00:00 Introduction
01:25 Vision Transformer Architectures
09:19 TinyCLIP's Hierarchy
13:21 Patch Sizes
14:32 Notebook 1: ViT Main Demo
17:27 Notebook 1: Direct Logit Attribution
20:01 Notebook 1: Question on Patch-Level Information and Emergent Segmentation Maps
27:23 Notebook 1 : Cat-Dog Switch
33:36 Notebook II: Emoji Logit Lens (Interactive)
38:27 Spider Man Emoji Logit Lens
40:51 Shoggoth, House, and More Emoji Logit Lenses
44:38 Notebook III: Attention Head Visualization

Twitter: https://x.com/soniajoseph_