Patrick Leask patrick.leask@durham.ac.uk
PGR Student Doctor of Philosophy
Sparse Autoencoders Do Not Find Canonical Units of Analysis
Leask, Patrick; Bussmann, Bart; Pearce, Michael T; Isaac Bloom, Joseph; Tigges, Curt; Al Moubayed, Noura; Sharkey, Lee; Nanda, Neel
Authors
Bart Bussmann
Michael T Pearce
Joseph Isaac Bloom
Curt Tigges
Dr Noura Al Moubayed noura.al-moubayed@durham.ac.uk
Associate Professor
Lee Sharkey
Neel Nanda
Abstract
A common goal of mechanistic interpretability is to decompose the activations of neural networks into features: interpretable properties of the input computed by the model. Sparse autoencoders (SAEs) are a popular method for finding these features in LLMs, and it has been postulated that they can be used to find a \textit{canonical} set of units: a unique and complete list of atomic features. We cast doubt on this belief using two novel techniques: SAE stitching to show they are incomplete, and meta-SAEs to show they are not atomic. SAE stitching involves inserting or swapping latents from a larger SAE into a smaller one. Latents from the larger SAE can be divided into two categories: novel latents, which improve performance when added to the smaller SAE, indicating they capture novel information, and \emph{reconstruction latents}, which can replace corresponding latents in the smaller SAE that have similar behavior. The existence of novel features indicates incompleteness of smaller SAEs. Using meta-SAEs - SAEs trained on the decoder matrix of another SAE - we find that latents in SAEs often decompose into combinations of latents from a smaller SAE, showing that larger SAE latents are not atomic. The resulting decompositions are often interpretable; e.g. a latent representing "Einstein'" decomposes into "scientist", "Germany", and "famous person". To train meta-SAEs we introduce BatchTopK SAEs, an improved variant of the popular TopK SAE method, that only enforces a fixed average sparsity. Even if SAEs do not find canonical units of analysis, they may still be useful tools. We suggest that future research should either pursue different approaches for identifying such units, or pragmatically choose the SAE size suited to their task. We provide an interactive dashboard to explore meta-SAEs: https://metasaes.streamlit.app/
Citation
Leask, P., Bussmann, B., Pearce, M. T., Isaac Bloom, J., Tigges, C., Al Moubayed, N., Sharkey, L., & Nanda, N. (2025, April). Sparse Autoencoders Do Not Find Canonical Units of Analysis. Presented at The Thirteenth International Conference on Learning Representations, Singapore
Presentation Conference Type | Conference Paper (published) |
---|---|
Conference Name | The Thirteenth International Conference on Learning Representations |
Start Date | Apr 24, 2025 |
End Date | Apr 26, 2025 |
Acceptance Date | Jan 22, 2025 |
Online Publication Date | Jan 22, 2025 |
Publication Date | Jan 22, 2025 |
Deposit Date | Feb 10, 2025 |
Publicly Available Date | Feb 12, 2025 |
Peer Reviewed | Peer Reviewed |
Keywords | sparse autoencoders, mechanistic interpretability |
Public URL | https://durham-repository.worktribe.com/output/3475474 |
Publisher URL | https://openreview.net/forum?id=9ca9eHNrdH |
Files
Published Conference Paper
(2.9 Mb)
PDF
Publisher Licence URL
http://creativecommons.org/licenses/by/4.0/
You might also like
Explainable text-tabular models for predicting mortality risk in companion animals
(2024)
Journal Article