Towards Compositionality in Concept Learning

Adam Stein, Aaditya Naik, Yinjun Wu, Mayur Naik, Eric Wong
Proceedings of the 41st International Conference on Machine Learning, PMLR 235:46530-46555, 2024.

Abstract

Concept-based interpretability methods offer a lens into the internals of foundation models by decomposing their embeddings into high-level concepts. These concept representations are most useful when they are compositional, meaning that the individual concepts compose to explain the full sample. We show that existing unsupervised concept extraction methods find concepts which are not compositional. To automatically discover compositional concept representations, we identify two salient properties of such representations, and propose Compositional Concept Extraction (CCE) for finding concepts which obey these properties. We evaluate CCE on five different datasets over image and text data. Our evaluation shows that CCE finds more compositional concept representations than baselines and yields better accuracy on four downstream classification tasks.

Cite this Paper

BibTeX
@InProceedings{pmlr-v235-stein24b, title = {Towards Compositionality in Concept Learning}, author = {Stein, Adam and Naik, Aaditya and Wu, Yinjun and Naik, Mayur and Wong, Eric}, booktitle = {Proceedings of the 41st International Conference on Machine Learning}, pages = {46530--46555}, year = {2024}, editor = {Salakhutdinov, Ruslan and Kolter, Zico and Heller, Katherine and Weller, Adrian and Oliver, Nuria and Scarlett, Jonathan and Berkenkamp, Felix}, volume = {235}, series = {Proceedings of Machine Learning Research}, month = {21--27 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v235/main/assets/stein24b/stein24b.pdf}, url = {https://proceedings.mlr.press/v235/stein24b.html}, abstract = {Concept-based interpretability methods offer a lens into the internals of foundation models by decomposing their embeddings into high-level concepts. These concept representations are most useful when they are compositional, meaning that the individual concepts compose to explain the full sample. We show that existing unsupervised concept extraction methods find concepts which are not compositional. To automatically discover compositional concept representations, we identify two salient properties of such representations, and propose Compositional Concept Extraction (CCE) for finding concepts which obey these properties. We evaluate CCE on five different datasets over image and text data. Our evaluation shows that CCE finds more compositional concept representations than baselines and yields better accuracy on four downstream classification tasks.} }
Endnote
%0 Conference Paper %T Towards Compositionality in Concept Learning %A Adam Stein %A Aaditya Naik %A Yinjun Wu %A Mayur Naik %A Eric Wong %B Proceedings of the 41st International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2024 %E Ruslan Salakhutdinov %E Zico Kolter %E Katherine Heller %E Adrian Weller %E Nuria Oliver %E Jonathan Scarlett %E Felix Berkenkamp %F pmlr-v235-stein24b %I PMLR %P 46530--46555 %U https://proceedings.mlr.press/v235/stein24b.html %V 235 %X Concept-based interpretability methods offer a lens into the internals of foundation models by decomposing their embeddings into high-level concepts. These concept representations are most useful when they are compositional, meaning that the individual concepts compose to explain the full sample. We show that existing unsupervised concept extraction methods find concepts which are not compositional. To automatically discover compositional concept representations, we identify two salient properties of such representations, and propose Compositional Concept Extraction (CCE) for finding concepts which obey these properties. We evaluate CCE on five different datasets over image and text data. Our evaluation shows that CCE finds more compositional concept representations than baselines and yields better accuracy on four downstream classification tasks.
APA
Stein, A., Naik, A., Wu, Y., Naik, M. & Wong, E.. (2024). Towards Compositionality in Concept Learning. Proceedings of the 41st International Conference on Machine Learning, in Proceedings of Machine Learning Research 235:46530-46555 Available from https://proceedings.mlr.press/v235/stein24b.html.

Related Material