Prototype-based geometry of local object patches in the pigeon Mesopallium Ventrolaterale

In natural vision, objects are frequently partially occluded and thus perceived through local fragments, yet how such fragments are organized in neural representational space remains poorly understood. Using pigeons as a model, we combined behavior from a whole–part image categorization paradigm with population electrophysiological recordings from the mesopallium ventrolaterale (MVL) during passive viewing. We used CEBRA-time to embed neural time series into a low-dimensional space, defined category neural prototypes from population responses to whole images, and projected local patches into the same space to quantify their distances to the prototypes. This framework allowed systematic tests of how prototype distance relates to behavioral accuracy and patch semantics. In MVL, animal categories formed compact, prototype-centered clusters. Across patches, greater prototype distance predicted lower behavioral accuracy, and patches closer to the prototype were enriched for diagnostic semantic parts (e.g., heads and faces). By contrast, across multiple deep-network baselines, prototype distance neither reliably predicted behavior nor consistently reflected the semantic distribution of patches. Building on these baselines, we then introduced a local prototype learning mechanism in which category matching is governed by the proximity between local features and learned prototypes. The resulting network reproduced several MVL-consistent geometric signatures, including prototype-centered clustering for animal categories and a negative relationship between prototype distance and behavior. Together, these results support the hypothesis that incomplete-object representations in pigeon MVL are organized by a diagnostic-part-related prototype geometry, and suggest that this geometry provides actionable neural constraints for brain-inspired vision models.