Caption Bottleneck Models

Recent advances in deep learning, exemplified by architectures like ResNet, Vision Transformer, and CLIP, have led to unprecedented performance in visual recognition tasks. However, these models are often criticized for their lack of interpretability, especially in critical applications such as healthcare diagnostics and autonomous vehicles. Concept Bottleneck Models (CBMs) emerged as a promising approach by introducing human-understandable intermediate concepts, enabling inspection and intervention. Notable works include Bau et al.'s concept-based explanations and foundation model-driven concept proposals. Despite progress, these models rely heavily on predefined concept sets, which are costly to annotate and often incomplete. Moreover, the static nature of concepts leads to limited flexibility and potential biases. Additionally, information leakage—where unmodeled visual cues influence predictions—remains a persistent challenge, compromising explanation fidelity. Recent efforts in automatic concept discovery and open-vocabulary methods aim to address these issues but still depend on external concept vocabularies, which may not fully capture dataset-specific nuances. CaBM builds upon these foundations, integrating natural language generation and post-hoc analysis to overcome these limitations, advancing the field toward more scalable and faithful interpretability.

The core challenge in developing interpretable vision models lies in balancing accuracy, interpretability, and robustness against information leakage. Static concept sets, though interpretable, are expensive and inflexible, limiting scalability. Moreover, existing CBMs are vulnerable to visual feature bypass, where unmodeled cues influence predictions, undermining trustworthiness. Achieving a leakage-free, flexible, and high-performance model requires rethinking the concept representation and training paradigm. The difficulty is compounded by the need for automatic, dataset-specific concept discovery without manual annotation, while ensuring that the recognition process remains transparent and faithful. These issues hinder the deployment of reliable AI in real-world, high-stakes scenarios.

CaBM introduces several key innovations: 1) It replaces static concept vectors with free-form natural language captions generated by a frozen large multimodal model, removing the need for predefined concept sets. 2) Implements structured prompting and taxonomy censoring to generate descriptions that exclude class labels, preventing trivial lexical cues. 3) Trains a text-only classifier on these captions, ensuring recognition occurs solely in language space, thus inherently avoiding visual information leakage. 4) Post-hoc, it employs gradient×embedding attribution and semantic clustering to automatically identify meaningful, dataset-specific concepts from the trained classifier’s decision process. This design guarantees leakage resistance while maintaining interpretability and competitive accuracy.

• �� Multi-caption generation: Use a pretrained LMM with structured prompts to produce K diverse captions per image, varying sampling temperature for lexical diversity.
• �� Taxonomy censoring: Apply deterministic filters to remove explicit class names and taxonomic terms, ensuring descriptions reflect visual attributes.
• �� Text classifier training: Encode captions with a transformer (e.g., RoBERTa), train with cross-entropy loss to predict labels, and aggregate logits across captions at inference.
• �� Concept discovery: Use gradient×embedding attribution to identify salient tokens, propose candidate phrases via n-grams, score importance through erasure (masking), and cluster semantically similar phrases to form concepts.
• �� Evaluation: Validate accuracy on multiple datasets, compare with baselines, and analyze the interpretability of extracted concepts through human and automatic assessments.

Experiments encompass six datasets: CIFAR-10, CIFAR-100, CUB-200, Food-101, Flowers-102, and ImageNet-1K. For each, five captions are generated per image at various temperatures, trained with a frozen LMM and RoBERTa-based classifier. The evaluation metrics include Top-1 accuracy, concept relevance, and interpretability. Results show CaBM achieves 78.5% Top-1 accuracy on ImageNet, comparable to state-of-the-art models, while automatically extracting human-understandable concepts. Ablation studies demonstrate the importance of structured prompting, taxonomy censoring, and multi-caption inference. The concept extraction pipeline produces semantically meaningful phrases, validated through human judgment and quantitative metrics, confirming the model’s interpretability and robustness against leakage.

CaBM attains competitive accuracy across datasets, e.g., 78.5% Top-1 on ImageNet, matching or surpassing existing CBMs. The automatically discovered concepts—such as 'reddish-orange crown' or 'black and white wings'—align with human intuition, providing faithful explanations. The leakage-resistant design prevents visual cues from bypassing the concept bottleneck, confirmed through ablation and adversarial tests. Furthermore, the model demonstrates strong generalization in fine- and coarse-grained tasks, with the ability to adapt to dataset-specific concepts without manual labeling or external dictionaries. These results validate the effectiveness of the approach in balancing interpretability and performance.