Abstract
Color is the prime example of categorical perception, yet it is still unclear why and how color categories emerge. While color categories have a functional role in communication, prelinguistic infants already respond to colors categorically. Here, we address the emergence of color categories as a result of the general interaction with the visual world. Specifically, we asked whether color categories arise in a convolutional neural network (CNN) trained to recognize objects in natural images. Therefore, we replaced the classification layer of a CNN (Resnet-18) trained on ImageNet and evaluated its performance on various color classification tasks. In Experiment 1 we trained the novel output layer to classify colored words sampled from the hue spectrum (HSV). Systematically varying stimulus colors demonstrates that the network not only generalizes similar colors, but also that the borders between classes are largely invariant to the training colors. Relying on the notion that colors from the same category should be easier to generalize than colors from different categories, in Experiment 2, an evolutionary algorithm finds similar border locations. Finally, in Experiment 3, we investigated a potential functional role of the found borders. Manipulating colors and color contrast we find that even in cluttered color stimuli, classification can still rely on the same borders. The fact that a CNN classifying objects in natural images represents color in a categorical manner, highlights that color categorization may emerge naturally with the development of basic visual skills. Considering the relative ease with which one can inspect the activity of individual neurons in a CNN, compared to a biological system, the current findings also open up an exciting research avenue for uncovering how color categories can be based on lower-level signals.