Self-learn to Explain Siamese Networks Robustly

TL;DR

This paper introduces a self-learning framework to improve the stability and faithfulness of gradient-based explanations for Siamese networks, especially under data scarcity and imbalance, with proven convergence and practical case studies.

Contribution

It proposes a novel optimization approach that enhances explanation stability for Siamese networks using self-learning and invariance, with convergence guarantees and empirical validation.

Findings

Framework improves explanation stability and faithfulness.

Convergence of optimization algorithms is experimentally validated.

Effective on tabular and graph data in neuroscience and chemical engineering.

Abstract

Learning to compare two objects are essential in applications, such as digital forensics, face recognition, and brain network analysis, especially when labeled data is scarce and imbalanced. As these applications make high-stake decisions and involve societal values like fairness and transparency, it is critical to explain the learned models. We aim to study post-hoc explanations of Siamese networks (SN) widely used in learning to compare. We characterize the instability of gradient-based explanations due to the additional compared object in SN, in contrast to architectures with a single input instance. We propose an optimization framework that derives global invariance from unlabeled data using self-learning to promote the stability of local explanations tailored for specific query-reference pairs. The optimization problems can be solved using gradient descent-ascent (GDA) for constrained optimization, or SGD for KL-divergence regularized unconstrained optimization, with convergence proofs, especially when the objective functions are nonconvex due to the Siamese architecture. Quantitative results and case studies on tabular and graph data from neuroscience and chemical engineering show that the framework respects the self-learned invariance while robustly optimizing the faithfulness and simplicity of the explanation. We further demonstrate the convergence of GDA experimentally.