关键词: 概念瓶颈模型, 变分信息瓶颈, 特征归因, 可解释性

Abstract: Concept bottleneck models (CBMs) project visual features extracted from black-box models onto a set of human-interpretable concepts to facilitate decision-making. Existing approaches typically rely on large language models (LLMs) to generate textual concepts and on multimodal pretrained models to align visual features with text embeddings. However, these methods often introduce textual noise into the bottleneck, resulting in explanations that may not accurately reflect the image content or its visual attributes. To address this limitation, a variational information bottleneck-guided complementary concept bottleneck model is proposed. This method employs a chain-of-thought (CoT)-based concept generation strategy that prompts both vision-language models (VLMs) and LLMs to produce more precise and complementary textual descriptions. A concept selection module based on variational information bottleneck feature attribution method is then developed to extract the textual concepts most relevant to the image content. Furthermore, an image classification strategy is designed that integrates dual-branch concept activation scores from complementary concept bottlenecks to support robust decision-making. Finally, an interpretability efficiency metric is introduced to evaluate the succinctness and effectiveness of the generated explanations. Experimental results on six public datasets demonstrate that the proposed method not only outperforms five state-of-the-art models in interpretability efficiency, but also achieves comparable or even superior classification accuracy.

Key words: concept bottleneck models, variational information bottleneck, feature attribution, interpretability