This work was produced as part of Lee Sharkey's stream in the ML Alignment & Theory Scholars Program - Winter 2023-24 Cohort

Intro and Motivation

Sparse dictionary learning (SDL) has attracted a lot of attention recently as a method for interpreting transformer activations. They demonstrate that model activations can often be explained using a sparsely-activating, overcomplete set of human-interpretable directions.

However, despite its success for explaining many components, applying SDL to interpretability is relatively nascent and have yet to be applied to some model activations. In particular, intermediate activations of attention blocks have yet to be studied, and provide challenges for standard SDL methods.

The first challenge is bilinearity: SDL is usually applied to individual vector spaces at individual layers, so we can simply identify features as a direction in activation space. But the QK circuits of transformer attention layers are different: They involve a bilinear [...]

---

Outline:

(00:16) Intro and Motivation

(02:09) Training Setup

(02:27) Step 1: Reconstructing the attention pattern with key- and query-transcoders

(02:36) Architecture

(03:25) Loss functions

(05:10) Step 2: Reducing to Sparse Feature-Pairs with Masking

(09:31) Results

(09:34) Both features and feature pairs are highly sparse

(10:17) Reconstructed attention patterns are highly accurate

(13:35) Feature Analysis

(13:55) Our unsupervised method identifies Name-Attention features in Name-Mover and Negative Name-Mover Heads

(17:18) Discovering Novel Feature-Pairs

(17:51) Example 1. Pushy Social Media (Layer 10)

(19:06) Example 2: Date Completion (Layer 10) - Attending from months to numbers which may be the day

(20:08) Feature Sparsity

(21:35) Key- and query-features activate densely

(22:45) A dense ‘Attend to BOS’ feature

(24:41) Discussion

(27:25) Future Work

The original text contained 5 footnotes which were omitted from this narration.

The original text contained 19 images which were described by AI.

---

First published:

Source:

Narrated by TYPE III AUDIO.