Archetypal SAE: Adaptive and Stable Dictionary Learning for Concept Extraction in Large Vision Models

Thomas Fel, Ekdeep Singh Lubana, Jacob S. Prince, Matthew Kowal, Victor Boutin, Isabel Papadimitriou, Binxu Wang, Martin Wattenberg, Demba E. Ba, Talia Konkle
Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:16543-16572, 2025.

Abstract

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: 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 data’s convex hull. This geometric anchoring significantly enhances the stability and plausibility 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.

Cite this Paper

BibTeX
@InProceedings{pmlr-v267-fel25a, title = {Archetypal {SAE}: Adaptive and Stable Dictionary Learning for Concept Extraction in Large Vision Models}, author = {Fel, Thomas and Lubana, Ekdeep Singh and Prince, Jacob S. and Kowal, Matthew and Boutin, Victor and Papadimitriou, Isabel and Wang, Binxu and Wattenberg, Martin and Ba, Demba E. and Konkle, Talia}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {16543--16572}, year = {2025}, editor = {Singh, Aarti and Fazel, Maryam and Hsu, Daniel and Lacoste-Julien, Simon and Berkenkamp, Felix and Maharaj, Tegan and Wagstaff, Kiri and Zhu, Jerry}, volume = {267}, series = {Proceedings of Machine Learning Research}, month = {13--19 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v267/main/assets/fel25a/fel25a.pdf}, url = {https://proceedings.mlr.press/v267/fel25a.html}, abstract = {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: 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 data’s convex hull. This geometric anchoring significantly enhances the stability and plausibility 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.} }
Endnote
%0 Conference Paper %T Archetypal SAE: Adaptive and Stable Dictionary Learning for Concept Extraction in Large Vision Models %A Thomas Fel %A Ekdeep Singh Lubana %A Jacob S. Prince %A Matthew Kowal %A Victor Boutin %A Isabel Papadimitriou %A Binxu Wang %A Martin Wattenberg %A Demba E. Ba %A Talia Konkle %B Proceedings of the 42nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2025 %E Aarti Singh %E Maryam Fazel %E Daniel Hsu %E Simon Lacoste-Julien %E Felix Berkenkamp %E Tegan Maharaj %E Kiri Wagstaff %E Jerry Zhu %F pmlr-v267-fel25a %I PMLR %P 16543--16572 %U https://proceedings.mlr.press/v267/fel25a.html %V 267 %X 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: 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 data’s convex hull. This geometric anchoring significantly enhances the stability and plausibility 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.
APA
Fel, T., Lubana, E.S., Prince, J.S., Kowal, M., Boutin, V., Papadimitriou, I., Wang, B., Wattenberg, M., Ba, D.E. & Konkle, T.. (2025). Archetypal SAE: Adaptive and Stable Dictionary Learning for Concept Extraction in Large Vision Models. Proceedings of the 42nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 267:16543-16572 Available from https://proceedings.mlr.press/v267/fel25a.html.

Related Material