Towards Cognitively Plausible Concept Learning: Spatially Grounding Concepts with Anatomical Priors

Understanding the internal cognitive processes of deep learning models is a critical challenge. Concept Bottleneck Models (CBMs) offer a path towards interpretability by mapping predictions to human-understandable concepts. However, we argue they suffer from a fundamental cognitive flaw: a "spatial grounding failure." Due to global pooling, standard CBMs are unable to connect concepts to their corresponding locations in an image, leading to activations in biologically implausible regions. This disconnect undermines their claim to providing a faithful processing account of their decisions. To address this, we introduce GroundedCBM, a framework inspired by how biological systems leverage structural priors for recognition. Our model embeds anatomical domain knowledge into the learning process through two core innovations: (1) a spatially-aware attention module that forces concepts to be localized in plausible regions, akin to how an expert uses an anatomical schema, and (2) a dynamic graph network that models contextual relationships between concepts, mimicking associative reasoning. On CUB-200-2011, GroundedCBM not only improves concept accuracy but also closes over 60\% of the performance gap to an equivalent black-box model. Our work demonstrates that by enforcing cognitively plausible spatial constraints, we can build models that provide a more faithful processing account of their cognition without sacrificing performance.