H-SPLID: HSIC-based Saliency Preserving Latent Information Decomposition

TL;DR

H-SPLID is a new algorithm that decomposes features into salient and non-salient parts, promoting low-dimensional, task-relevant representations and linking robustness to feature compression, demonstrated on image classification tasks.

Contribution

It introduces H-SPLID, a novel method for explicit feature decomposition that enhances salient feature learning and robustness in neural networks.

Findings

Models with H-SPLID rely more on salient features.

H-SPLID reduces sensitivity to background perturbations.

Theoretical bounds relate robustness to feature subspace dimensions.

Abstract

We introduce H-SPLID, a novel algorithm for learning salient feature representations through the explicit decomposition of salient and non-salient features into separate spaces. We show that H-SPLID promotes learning low-dimensional, task-relevant features. We prove that the expected prediction deviation under input perturbations is upper-bounded by the dimension of the salient subspace and the Hilbert-Schmidt Independence Criterion (HSIC) between inputs and representations. This establishes a link between robustness and latent representation compression in terms of the dimensionality and information preserved. Empirical evaluations on image classification tasks show that models trained with H-SPLID primarily rely on salient input components, as indicated by reduced sensitivity to perturbations affecting non-salient features, such as image backgrounds. Our code is available at https://github.com/neu-spiral/H-SPLID.