Concepts from Representations: Post-hoc Concept Bottleneck Models via Sparse Decomposition of Visual Representations

TL;DR

This paper tackles the opacity of deep image classifiers by introducing PCBM-ReD, a post-hoc pipeline that harnesses data-driven concept discovery, multimodal labeling, and reconstruction-guided selection to build an independent, sparse concept bottleneck that decomposes a pretrained encoder's representations. By labeling concepts with MLLMs, ensuring visual identifiability, and leveraging CLIP's visual-text alignment, PCBM-ReD reconstructs image embeddings as sparse linear combinations of concept embeddings and trains a simple linear head, yielding competitive accuracy with enhanced interpretability. The approach demonstrates strong performance across 11 datasets, including zero-shot and few-shot scenarios, and outperforms several CBMs without manual concept curation. The work provides a scalable path to interpretable, high-performing models in diverse domains, while acknowledging limitations related to domain-specific knowledge in MLLMs and dependence on the pretrained encoder.

Abstract

Deep learning has achieved remarkable success in image recognition, yet their inherent opacity poses challenges for deployment in critical domains. Concept-based interpretations aim to address this by explaining model reasoning through human-understandable concepts. However, existing post-hoc methods and ante-hoc concept bottleneck models (CBMs), suffer from limitations such as unreliable concept relevance, non-visual or labor-intensive concept definitions, and model or data-agnostic assumptions. This paper introduces Post-hoc Concept Bottleneck Model via Representation Decomposition (PCBM-ReD), a novel pipeline that retrofits interpretability onto pretrained opaque models. PCBM-ReD automatically extracts visual concepts from a pre-trained encoder, employs multimodal large language models (MLLMs) to label and filter concepts based on visual identifiability and task relevance, and selects an independent subset via reconstruction-guided optimization. Leveraging CLIP's visual-text alignment, it decomposes image representations into linear combination of concept embeddings to fit into the CBMs abstraction. Extensive experiments across 11 image classification tasks show PCBM-ReD achieves state-of-the-art accuracy, narrows the performance gap with end-to-end models, and exhibits better interpretability.

Figures (12)

• Figure 1: PCBM-ReD extracts concepts from the pre-trained image encoder and reconstruct the visual representation with the concepts, which gives faithful interpretation and can take the best advantage of the encoder's representation power.
• Figure 2: We present an overview of Post-hoc Concept Bottleneck Model via Representation Decomposition (PCBM-ReD). First, we extract concepts from the learned representation, and use MLLMs to summarize and score the concepts. Second, we apply a concept selection algorithm to choose concepts and construct the bottleneck. Third, we perform sparse decomposition and reconstruct the image embedding by concepts. A linear layer is trained to predict the targets with the fitted embedding.
• Figure 3: Few-shot test accuracy (%) comparison. Average test accuracy on 11 datasets is reported. Shot means the number of labeled images for each class.
• Figure 4: Several example explanations generated by PCBM-ReD. The examples are sampled from the test set of 11 datasets, which have correct predictions. We also show their corresponding top concepts that contribute the most to the logits.
• Figure 5: Human evaluation. Volunteers rate the explanation on a scale of 1 to 5 (5 = very agree). S1: The explanations are visually identifiable features. S2: The explanations faithfully describe the image. S3: There is a causal relationship between the explanation and the prediction.