The representation of object-concept primitives across the macaque brain

Object concepts are mental representations that provide information beyond immediate sensory input, but where and how the brain encodes them remains unresolved. Here, macaques spent years learning a set of object-concept primitives: two-dimensional stimuli, each defined by a paired association between a unique color and a unique shape. During fMRI, the animals were cued with one feature (shapeless color or colorless shape) and rewarded for subsequently identifying the associated feature. Using a new whole-brain searchlight inspired by convolutional networks, we identified regions that carried information about the unshown feature during cue presentation. Cross-feature decoding was absent in retinotopic cortex but present bilaterally in discrete foci in anterior inferior temporal cortex (aIT), the temporal pole (TP), and frontal cortex (FC). In aIT and TP, the geometry of the neural representation of shapes matched the color-circle geometry of the associated colors, whereas FC supported the best decoding of error trials. Additional passive-viewing experiments showed univariate fMRI response biases to the learned associations in extrastriate visual cortex, which may reflect feedback from regions identified by cross-feature decoding. Together, the results provide a blueprint for how the brain encodes object concepts and suggest how concept representations engage perceptual representations through feedback during vision.