Evaluating the Interpretability of Sparse Autoencoders with Concept Annotations

The authors present a novel evaluation framework for Sparse Autoencoders (SAEs) that quantifies the semantic alignment between SAE latents and human-annotated concepts, using a new method called Fully-Binary Matching Pursuit (FBMP). They introduce synthetic benchmarks, synCUB and synCOCO, to facilitate targeted attribute perturbations and propose the Targeted Attribute Perturbation Alignment Score (TAPAScore) to assess the interpretability of SAEs trained on CLIP and DINOv2 embeddings. This framework enables practitioners to better evaluate and optimize SAEs for interpretability, suggesting that moderate dictionary sizes yield the best performance in aligning concepts with human understanding.