ResearchPod Summary
Concept Bottleneck Models (CBMs) aim to make AI reasoning transparent by forcing models to predict human-interpretable concepts before making a final classification. However, existing vision-language CBMs often struggle with fine-grained localization and rely on rigid, pre-aligned embedding spaces (like CLIP) that do not account for the true semantic geometry between visual patches and textual concepts. This paper asks: how can we move beyond static projection to create a dynamic, learnable, and spatially grounded cross-modal alignment for more faithful concept reasoning?
The authors propose the Optimal Transport Flow Concept Bottleneck Model (OTF-CBM). Instead of using fixed cosine similarity to align images and text, the model treats alignment as an optimal transport problem. It employs two key innovations:
OTF-CBM provides a more flexible and accurate framework for concept-based reasoning. By explicitly modeling the transport geometry, the model achieves superior classification accuracy and higher concept faithfulness compared to state-of-the-art CBM variants. The velocity-based activation mechanism allows the model to capture interpretable geometric relations without the computational overhead of integrating ordinary differential equations (ODEs) at inference time.
This work shifts the paradigm of interpretable AI from static, rigid mapping to a dynamic, geometric process. By enabling models to learn their own cross-modal metrics, OTF-CBM bridges the gap between raw visual evidence and human-understandable concepts, making AI systems more transparent, robust, and capable of fine-grained reasoning in complex vision-language tasks.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.