Identifying Sparsely Active Circuits Through Local Loss Landscape Decomposition

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Chrisman, Brianna, Bushnaq, Lucius, Sharkey, Lee
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908292492558336
author Chrisman, Brianna
Bushnaq, Lucius
Sharkey, Lee
author_facet Chrisman, Brianna
Bushnaq, Lucius
Sharkey, Lee
contents Much of mechanistic interpretability has focused on understanding the activation spaces of large neural networks. However, activation space-based approaches reveal little about the underlying circuitry used to compute features. To better understand the circuits employed by models, we introduce a new decomposition method called Local Loss Landscape Decomposition (L3D). L3D identifies a set of low-rank subnetworks: directions in parameter space of which a subset can reconstruct the gradient of the loss between any sample's output and a reference output vector. We design a series of progressively more challenging toy models with well-defined subnetworks and show that L3D can nearly perfectly recover the associated subnetworks. Additionally, we investigate the extent to which perturbing the model in the direction of a given subnetwork affects only the relevant subset of samples. Finally, we apply L3D to a real-world transformer model and a convolutional neural network, demonstrating its potential to identify interpretable and relevant circuits in parameter space.
format Preprint
id arxiv_https___arxiv_org_abs_2504_00194
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Identifying Sparsely Active Circuits Through Local Loss Landscape Decomposition
Chrisman, Brianna
Bushnaq, Lucius
Sharkey, Lee
Machine Learning
Artificial Intelligence
Much of mechanistic interpretability has focused on understanding the activation spaces of large neural networks. However, activation space-based approaches reveal little about the underlying circuitry used to compute features. To better understand the circuits employed by models, we introduce a new decomposition method called Local Loss Landscape Decomposition (L3D). L3D identifies a set of low-rank subnetworks: directions in parameter space of which a subset can reconstruct the gradient of the loss between any sample's output and a reference output vector. We design a series of progressively more challenging toy models with well-defined subnetworks and show that L3D can nearly perfectly recover the associated subnetworks. Additionally, we investigate the extent to which perturbing the model in the direction of a given subnetwork affects only the relevant subset of samples. Finally, we apply L3D to a real-world transformer model and a convolutional neural network, demonstrating its potential to identify interpretable and relevant circuits in parameter space.
title Identifying Sparsely Active Circuits Through Local Loss Landscape Decomposition
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2504.00194