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

Join us for an insightful talk on Goodfire's AdVersarial Parameter Decomposition Technique!

Most interpretability work focuses on decomposing a model's activations, but Goodfire's innovative approach breaks down a network's weights into components that serve distinct computational roles, such as predicting emoticons and identifying gender.

This method transforms the model from a simple lookup table into a generalizing algorithm, effectively handling attention computations across multiple heads. It stands as a competitor to SAEs and transcoders, allowing researchers to manually edit model behavior for interpretable and predictable changes.

Speaker: Linda Linsefors, co-author of Goodfire's Interpreting Language Model Parameters paper, will guide us through the workings of AdVersarial Parameter Decomposition (VPD), its findings, and future directions.

🔗 Research here

Schedule:

• 🕖 7:00 PM — Doors open, snacks and refreshments served
• 🕢 7:30 PM — Talk begins!
• 🕣 8:30 PM — Wrap up Q&A, continued hangouts

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