Faithful and Plausible Natural Language Explanations for Image Classification: A Pipeline Approach

TL;DR

This paper introduces a post-hoc natural language explanation pipeline for CNN image classifiers that produces faithful, plausible, and grounded explanations without affecting the classifier's performance.

Contribution

It presents a novel pipeline approach that generates faithful and plausible natural language explanations by analyzing influential neurons and activation maps, applicable to any CNN classifier.

Findings

Generated explanations are more faithful and plausible than baselines.

User interventions improve explanation quality threefold.

Method does not alter classifier training or performance.

Abstract

Existing explanation methods for image classification struggle to provide faithful and plausible explanations. This paper addresses this issue by proposing a post-hoc natural language explanation method that can be applied to any CNN-based classifier without altering its training process or affecting predictive performance. By analysing influential neurons and the corresponding activation maps, the method generates a faithful description of the classifier's decision process in the form of a structured meaning representation, which is then converted into text by a language model. Through this pipeline approach, the generated explanations are grounded in the neural network architecture, providing accurate insight into the classification process while remaining accessible to non-experts. Experimental results show that the NLEs constructed by our method are significantly more plausible and faithful. In particular, user interventions in the neural network structure (masking of neurons) are three times more effective than the baselines.