Saved in:
Bibliographic Details
Main Authors: Gross, Jason, Agrawal, Rajashree, Kwa, Thomas, Ong, Euan, Yip, Chun Hei, Gibson, Alex, Noubir, Soufiane, Chan, Lawrence
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2406.11779
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916540909092864
author Gross, Jason
Agrawal, Rajashree
Kwa, Thomas
Ong, Euan
Yip, Chun Hei
Gibson, Alex
Noubir, Soufiane
Chan, Lawrence
author_facet Gross, Jason
Agrawal, Rajashree
Kwa, Thomas
Ong, Euan
Yip, Chun Hei
Gibson, Alex
Noubir, Soufiane
Chan, Lawrence
contents We propose using mechanistic interpretability -- techniques for reverse engineering model weights into human-interpretable algorithms -- to derive and compactly prove formal guarantees on model performance. We prototype this approach by formally proving accuracy lower bounds for a small transformer trained on Max-of-K, validating proof transferability across 151 random seeds and four values of K. We create 102 different computer-assisted proof strategies and assess their length and tightness of bound on each of our models. Using quantitative metrics, we find that shorter proofs seem to require and provide more mechanistic understanding. Moreover, we find that more faithful mechanistic understanding leads to tighter performance bounds. We confirm these connections by qualitatively examining a subset of our proofs. Finally, we identify compounding structureless errors as a key challenge for using mechanistic interpretability to generate compact proofs on model performance.
format Preprint
id arxiv_https___arxiv_org_abs_2406_11779
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Compact Proofs of Model Performance via Mechanistic Interpretability
Gross, Jason
Agrawal, Rajashree
Kwa, Thomas
Ong, Euan
Yip, Chun Hei
Gibson, Alex
Noubir, Soufiane
Chan, Lawrence
Machine Learning
Logic in Computer Science
We propose using mechanistic interpretability -- techniques for reverse engineering model weights into human-interpretable algorithms -- to derive and compactly prove formal guarantees on model performance. We prototype this approach by formally proving accuracy lower bounds for a small transformer trained on Max-of-K, validating proof transferability across 151 random seeds and four values of K. We create 102 different computer-assisted proof strategies and assess their length and tightness of bound on each of our models. Using quantitative metrics, we find that shorter proofs seem to require and provide more mechanistic understanding. Moreover, we find that more faithful mechanistic understanding leads to tighter performance bounds. We confirm these connections by qualitatively examining a subset of our proofs. Finally, we identify compounding structureless errors as a key challenge for using mechanistic interpretability to generate compact proofs on model performance.
title Compact Proofs of Model Performance via Mechanistic Interpretability
topic Machine Learning
Logic in Computer Science
url https://arxiv.org/abs/2406.11779