Lina Teichmann¹, Ka Chun Lam¹, Amaia Benitez-Andonegui¹, Danny Garside^2,3, Sebastian Montesinos¹, Bin Guan², Francisco Pereira¹, Chris I. Baker¹, Bevil R. Conway²; ¹National Institute of Mental Health, ²National Eye Institute, ³University of Sussex

Color perception depends on the ability to extract differences in the spectral composition of light. In humans with normal color vision this process is initiated by the absorption of light by three classes of cone photoreceptors (L, M, S) followed by cone-opponent operations carried out by bipolar cells. But these color encoding mechanisms do not fully predict color appearance. Many attempts have been made to construct formal color spaces that capture uniform perceptual similarity relationships among colors, but discrepancies remain between the models and individual perception. Combining Magnetoencephalography (MEG) and psychophysical data obtained in the same set of participants, we examined the extent to which these discrepancies and individual differences can be accounted for by the geometry of the neural representation of color and its evolution over time. Unlike previous work, we collected neural responses from eight individuals who each viewed hundreds of unique colors presented in 16,200 color trials. The dense sampling of color space allowed us to reconstruct a fine-grained geometry of the neural representation of color in individual participants, with millisecond accuracy. We related these neurophysiological measurements to similarity judgements collected in the same participants. Using Variational Interpretable Concept Embeddings, we extracted similarity embeddings from the behavioral data to model the neural signal directly using regression models. Consistent with published results, we find that color information is present in the neural signal from approximately 70 ms onwards. Importantly, we discovered that the neural color-space geometries unfolded non-uniformly over time. In particular, the representational structures read out with MEG are best characterized across all participants by models of cone-opponent encoding mechanisms, especially early in the timeseries. These findings highlight the gap between theoretical color spaces and color perception and provide a novel avenue to gain insights into the subjective nature of perception.