Abstract

Object recognition requires mapping variable sensory inputs to stable object identities, defining how input variations preserve identity or represent related versus unrelated objects. These mappings often generalize rapidly to novel examples under minimal learning. To study how the brain uses generalizable structure for odor object recognition, we used optogenetically-driven “synthetic odors”—allowing precise, parametric variations—while measuring perceptual responses in mice, and separately, recording from odor object cortex. We found that odor recognition was best explained as the matching of spatio-temporally defined features in input activity. Additionally, cortical activity in naïve animals contained structured representations that precisely anticipated perceptual judgments in trained animals. This structure was absent in randomly-initialized recurrent network models and input patterns themselves, and was best described by low-dimensional hyperspheric geometry, ideal for rapid learning and generalization. These results reveal a pre-existing neural scaffold for object recognition that supports flexible and generalizable learning. Our synthetic approach reveals the fundamental logic of the olfactory code and provides a general framework for testing links between sensory representations and perception.