C2F-Explainer: Explaining Transformers Better Through a Coarse-to-Fine Strategy

Abstract

Transformer interpretability research is a hot topic in the area of deep learning. Traditional interpretation methods mostly use the final layer output of the Transformer encoder as masks to generate an explanation map. However, These approaches overlook two crucial aspects. At the coarse-grained level, the mask may contain uncertain information, including unreliable and incomplete object location data; at the fine-grained level, there is information loss on the mask, resulting in spatial noise and detail loss. To address these issues, in this paper, we propose a two-stage coarse-to-fine strategy (C2F-Explainer) for improving Transformer interpretability. Specifically, we first design a sequential three-way mask (S3WM) module to handle the problem of uncertain information at the coarse-grained level. This module uses sequential three-way decisions to process the mask, preventing uncertain information on the mask from impacting the interpretation results, thus obtaining coarse-grained interpretation results with accurate position. Second, to further reduce the impact of information loss at the fine-grained level, we devised an attention fusion (AF) module inspired by the fact that self-attention can capture global semantic information, AF aggregates the attention matrix to generate a cross-layer relation matrix, which is then used to optimize detailed information on the interpretation results and produce fine-grained interpretation results with clear and complete edges. Experimental results show that the proposed C2F-Explainer has good interpretation results on both natural and medical image datasets, and the mIoU is improved by 2.08% on the PASCAL VOC 2012 dataset.