Dynamic Clue Bottlenecks: Towards Interpretable-by-Design Visual Question Answering

TL;DR

This paper introduces DCLUB, an inherently interpretable visual question answering model that provides human-readable visual clues as intermediate explanations, improving reasoning accuracy while maintaining high performance.

Contribution

The paper proposes the DCLUB model, which offers an interpretable VQA system with intermediate visual clues, and introduces a new dataset for training and evaluating explanations.

Findings

DCLUB improves reasoning accuracy by 4.64% over black-box models.

DCLUB maintains 99.43% of VQA-v2 performance.

The model provides human-readable visual clues as explanations.

Abstract

Recent advances in multimodal large language models (LLMs) have shown extreme effectiveness in visual question answering (VQA). However, the design nature of these end-to-end models prevents them from being interpretable to humans, undermining trust and applicability in critical domains. While post-hoc rationales offer certain insight into understanding model behavior, these explanations are not guaranteed to be faithful to the model. In this paper, we address these shortcomings by introducing an interpretable by design model that factors model decisions into intermediate human-legible explanations, and allows people to easily understand why a model fails or succeeds. We propose the Dynamic Clue Bottleneck Model ( (DCLUB), a method that is designed towards an inherently interpretable VQA system. DCLUB provides an explainable intermediate space before the VQA decision and is faithful from the beginning, while maintaining comparable performance to black-box systems. Given a question, DCLUB first returns a set of visual clues: natural language statements of visually salient evidence from the image, and then generates the output based solely on the visual clues. To supervise and evaluate the generation of VQA explanations within DCLUB, we collect a dataset of 1.7k reasoning-focused questions with visual clues. Evaluations show that our inherently interpretable system can improve 4.64% over a comparable black-box system in reasoning-focused questions while preserving 99.43% of performance on VQA-v2.