How to take apart a neural network (or, Goodfire's AdVersarial Parameter Decomposition Technique)

​Most interpretability work decomposes a model's activations; Goodfire's new approach decomposes its weights, breaking a network's parameters into components that each play a distinct computational role: predicting emoticons, identifying gender, and more.

The result behaves less like a lookup table and more like a generalizing algorithm, and it's the first method to satisfyingly handle attention computations, even when they're spread across multiple heads. It's now a competitor to SAEs and transcoders, and lets researchers edit model behavior by hand, making interpretable, predictable changes.

​Linda Linsefors, co-author of Goodfire's Interpreting Language Model Parameters paper, will walk us through AdVersarial Parameter Decomposition (VPD): how it works, what it's found, and where it's headed.

​Research here: https://www.goodfire.ai/research/vpd-explainer#

​Schedule
7:00pm — doors open, snacks and refreshments served
7:30pm — talk begins!
8:30pm — wrap up Q&A, continued hangouts

​This event is a part of the Mox Summer Season: talks and socials at the frontier of ideas, running June through August.