Keuper Labs
← All publications

ICML 25 · 2025

DCBM - Data-Efficient Visual Concept Bottleneck Models

Katharina Prasse, Patrick Knab, Sascha Marton, Christian Bartelt, Margret Keuper

Explainable AIConcept Bottleneck ModelsFew-Shot Learning

Why this publication matters

An interpretable classifier should show which visual evidence supports its answer. DCBM learns a compact collection of concepts from image regions and can locate those concepts in new examples. It offers an efficient approach for specialized recognition tasks where both training data and understandable explanations matter.

Abstract

Concept Bottleneck Models (CBMs) enhance the interpretability of neural networks by basing predictions on human-understandable concepts. However, current CBMs typically rely on concept sets extracted from large language models or extensive image corpora, limiting their effectiveness in data-sparse scenarios. We propose Data-efficient CBMs (DCBMs), which reduce the need for large sample sizes during concept generation while preserving interpretability. DCBMs define concepts as image regions detected by segmentation or detection foundation models, allowing each image to generate multiple concepts across different granularities. Exclusively containing dataset-specific concepts, DCBMs are well suited for fine-grained classification and outof-distribution tasks. Attribution analysis using Grad-CAM demonstrates that DCBMs deliver visual concepts that can be localized in test images. By leveraging dataset-specific concepts instead of predefined or general ones, DCBMs enhance adaptability to new domains. The code is available at: https://github.com/KathPra/DCBM.

Abstract source ↗

Figures

Visual concepts extracted from training-image regions and used for an interpretable prediction.
Figure 1. DCBMs extract image regions as concepts. Using vision foundation models, we use crop image regions as concepts for CBM training. Based on few concept samples (50 imgs / class), DCBMs offer interpretability even for fine-grained classification. View in source ↗
DCBM framework.
Figure 2. DCBM framework. The DCBM framework generates concept proposals through foundation models (Step 1). These proposals are then clustered, each represented by its centroid (Step 2). Finally, the unique concepts are utilized to train a sparse CBM, effectively (Step 3). We leverage the image-text alignment to map the visual concept to the corresponding textual concept (Step 4). We can remove undesired concepts after Step 2. View in source ↗

Cite this paper

Download .bib 
@inproceedings{prasse2025dcbmdataefficient67,
  title = {{DCBM - Data-Efficient Visual Concept Bottleneck Models}},
  author = {Katharina Prasse and Patrick Knab and Sascha Marton and Christian Bartelt and Margret Keuper},
  booktitle = {International Conference on Machine Learning},
  year = {2025},
  url = {https://arxiv.org/pdf/2412.11576}
}

Figures and abstract are reproduced from the linked research sources. Credit remains with the authors and publishers.