Group Equivariance Meets Mechanistic Interpretability: Equivariant Sparse Autoencoders
This work addresses the problem of making mechanistic interpretability tools more effective for domains with symmetries, representing an incremental advancement by extending SAEs to handle group equivariance.
The paper tackled the challenge of adapting sparse autoencoders (SAEs) to domains with group symmetries, such as scientific data, by incorporating these symmetries into SAEs, resulting in features that improved downstream task performance, with adaptive equivariant SAEs showing superior probing performance compared to regular SAEs.
Sparse autoencoders (SAEs) have proven useful in disentangling the opaque activations of neural networks, primarily large language models, into sets of interpretable features. However, adapting them to domains beyond language, such as scientific data with group symmetries, introduces challenges that can hinder their effectiveness. We show that incorporating such group symmetries into the SAEs yields features more useful in downstream tasks. More specifically, we train autoencoders on synthetic images and find that a single matrix can explain how their activations transform as the images are rotated. Building on this, we develop adaptively equivariant SAEs that can adapt to the base model's level of equivariance. These adaptive SAEs discover features that lead to superior probing performance compared to regular SAEs, demonstrating the value of incorporating symmetries in mechanistic interpretability tools.