Decoding Dark Matter: Specialized Sparse Autoencoders for Interpreting Rare Concepts in Foundation Models

TL;DR

This paper introduces Specialized Sparse Autoencoders (SSAEs) that effectively interpret rare, crucial concepts in foundation models by focusing on specific subdomains, improving interpretability and bias mitigation.

Contribution

The paper presents SSAEs with a practical training recipe, demonstrating their ability to capture tail concepts and reduce spurious correlations in foundation models.

Findings

SSAEs outperform general-purpose SAEs in capturing subdomain tail concepts.

SSAEs increase worst-group classification accuracy by 12.5% on Bias in Bios dataset.

Tilted Empirical Risk Minimization improves concept recall.

Abstract

Understanding and mitigating the potential risks associated with foundation models (FMs) hinges on developing effective interpretability methods. Sparse Autoencoders (SAEs) have emerged as a promising tool for disentangling FM representations, but they struggle to capture rare, yet crucial concepts in the data. We introduce Specialized Sparse Autoencoders (SSAEs), designed to illuminate these elusive dark matter features by focusing on specific subdomains. We present a practical recipe for training SSAEs, demonstrating the efficacy of dense retrieval for data selection and the benefits of Tilted Empirical Risk Minimization as a training objective to improve concept recall. Our evaluation of SSAEs on standard metrics, such as downstream perplexity and $L_0$ sparsity, show that they effectively capture subdomain tail concepts, exceeding the capabilities of general-purpose SAEs. We showcase the practical utility of SSAEs in a case study on the Bias in Bios dataset, where SSAEs achieve a 12.5\% increase in worst-group classification accuracy when applied to remove spurious gender information. SSAEs provide a powerful new lens for peering into the inner workings of FMs in subdomains.