Is Sparse Attention more Interpretable?

TL;DR

This paper investigates whether sparse attention mechanisms genuinely enhance interpretability in models, finding that sparsity does not improve and may even hinder the ability to interpret model behavior through attention distributions.

Contribution

The study provides empirical evidence that sparsity in attention does not strengthen the interpretability of models and challenges assumptions about attention as an explainability tool.

Findings

Weak relationship between inputs and internal representations regardless of sparsity

Sparse attention does not map to influential inputs effectively

Sparsity may reduce the interpretability of attention distributions

Abstract

Sparse attention has been claimed to increase model interpretability under the assumption that it highlights influential inputs. Yet the attention distribution is typically over representations internal to the model rather than the inputs themselves, suggesting this assumption may not have merit. We build on the recent work exploring the interpretability of attention; we design a set of experiments to help us understand how sparsity affects our ability to use attention as an explainability tool. On three text classification tasks, we verify that only a weak relationship between inputs and co-indexed intermediate representations exists -- under sparse attention and otherwise. Further, we do not find any plausible mappings from sparse attention distributions to a sparse set of influential inputs through other avenues. Rather, we observe in this setting that inducing sparsity may make it less plausible that attention can be used as a tool for understanding model behavior.