BROWSE
Detail
Conference Paper
LoCo-Bot: Intrinsically Grounded Concept Bottleneck Models in a Single Forward Pass
Cited - time in 
Share 
Share
- Authors
- Sangwon Kim, Kyoungoh Lee, In-su Jang, Kwang-Ju Kim
- Issue Date
- 2026-09
- Citation
- European Conference on Computer Vision (ECCV) 2026 Workshop : Privacy, Fairness, Accountability and Transparency in Computer Vision (PFATCV), pp.1-8
- Publisher
- Privacy, Fairness, Accountability and Transparency in Computer Vision (PFATCV)
- Language
- English
- Type
- Conference Paper
- Abstract
- Concept Bottleneck Models (CBMs) expose interpretable concepts, yet rarely reveal where their visual evidence lies. Existing grounded CBMs often add an external grounding model or a separate localization pathway, so the displayed region need not be structurally tied to the computation producing the concept prediction. We introduce LoCo-Bot, an intrinsically grounded CBM that predicts concept scores and evidence maps from the same patch-level energy landscape in a single forward pass. Using only image-level concept labels, LoCo-Bot compares concept-specific patch features with positive and negative Concept Activation Vectors (CAVs). Soft multiple-instance pooling of these energies yields concept probabilities, while their patch-wise contrast defines structurally coupled spatial evidence. A dual CAV codebook further enables direct concept intervention. Across CUB-200-2011, AwA2, and CIFAR-100, LoCo-Bot provides competitive recognition and strong concept accuracy. On CUB, it attains 79.07% Inclusion and produces evidence whose removal decreases the corresponding concept score more than GradCAM, empirically supporting faithful grounding without external models or additional backward passes.
- Keyword
- Concept Bottleneck Models, Concept Localization, Interpretable Vision
- KSP Keywords
- CIFAR-10, Patch-wise, Positive and negative, Strong concept, Visual evidence, energy landscape, positive and