Decoding the yellow of a gray banana.

Some everyday objects are associated with a particular color, such as bananas, which are typically yellow. Behavioral studies show that perception of these so-called color-diagnostic objects is influenced by our knowledge of their typical color, referred to as memory color. However, neural representations of memory colors are unknown. Here we investigated whether memory color can be decoded from visual cortex activity when color-diagnostic objects are viewed as grayscale images. We trained linear classifiers to distinguish patterns of fMRI responses to four different hues. We found that activity in V1 allowed predicting the memory color of color-diagnostic objects presented in grayscale in naive participants performing a motion task. The results imply that higher areas feed back memory-color signals to V1. When classifiers were trained on neural responses to some exemplars of color-diagnostic objects and tested on others, areas V4 and LOC also predicted memory colors. Representational similarity analysis showed that memory-color representations in V1 were correlated specifically with patterns in V4 but not LOC. Our findings suggest that prior knowledge is projected from midlevel visual regions onto primary visual cortex, consistent with predictive coding theory.