Sum-of-Parts: Self-Attributing Neural Networks with End-to-End Learning of Feature Groups
This work addresses the problem of interpretable models for high-dimensional problems in machine learning, offering a novel method to improve performance without sacrificing interpretability, though it is incremental in advancing SANNs.
The paper tackles the performance trade-offs in self-attributing neural networks (SANNs) by proving a lower bound on errors for per-feature SANNs and showing that group-based SANNs can achieve zero error. It proposes Sum-of-Parts (SOP), a framework that transforms models into group-based SANNs with end-to-end learning of feature groups, achieving state-of-the-art performance on vision and language tasks and validating interpretability on quantitative and semantic metrics.
Self-attributing neural networks (SANNs) present a potential path towards interpretable models for high-dimensional problems, but often face significant trade-offs in performance. In this work, we formally prove a lower bound on errors of per-feature SANNs, whereas group-based SANNs can achieve zero error and thus high performance. Motivated by these insights, we propose Sum-of-Parts (SOP), a framework that transforms any differentiable model into a group-based SANN, where feature groups are learned end-to-end without group supervision. SOP achieves state-of-the-art performance for SANNs on vision and language tasks, and we validate that the groups are interpretable on a range of quantitative and semantic metrics. We further validate the utility of SOP explanations in model debugging and cosmological scientific discovery. Our code is available at https://github.com/BrachioLab/sop