Archetypal SAE: Adaptive and Stable Dictionary Learning for Concept Extraction in Large Vision Models
This addresses a fundamental reliability problem for researchers using SAEs to interpret large vision models, though it is an incremental improvement based on existing frameworks.
The paper tackled the instability of Sparse Autoencoders (SAEs) in machine learning interpretability by proposing Archetypal SAEs (A-SAE) and their variants (RA-SAEs), which constrain dictionary atoms to the convex hull of data, resulting in significantly enhanced stability and state-of-the-art reconstruction abilities.
Sparse Autoencoders (SAEs) have emerged as a powerful framework for machine learning interpretability, enabling the unsupervised decomposition of model representations into a dictionary of abstract, human-interpretable concepts. However, we reveal a fundamental limitation: existing SAEs exhibit severe instability, as identical models trained on similar datasets can produce sharply different dictionaries, undermining their reliability as an interpretability tool. To address this issue, we draw inspiration from the Archetypal Analysis framework introduced by Cutler & Breiman (1994) and present Archetypal SAEs (A-SAE), wherein dictionary atoms are constrained to the convex hull of data. This geometric anchoring significantly enhances the stability of inferred dictionaries, and their mildly relaxed variants RA-SAEs further match state-of-the-art reconstruction abilities. To rigorously assess dictionary quality learned by SAEs, we introduce two new benchmarks that test (i) plausibility, if dictionaries recover "true" classification directions and (ii) identifiability, if dictionaries disentangle synthetic concept mixtures. Across all evaluations, RA-SAEs consistently yield more structured representations while uncovering novel, semantically meaningful concepts in large-scale vision models.