Emergent Color Categorization in a Neural Network trained for Object Recognition

Color is a prime example of categorical perception, yet it is still unclear why and how color categories emerge. The key questions revolve around to what extent perceptual and linguistic processes shape categories. While prelinguistic infants and animals appear to treat color categorically, several recent attempts to model category formation have successfully utilized communicative concepts to predict color categories. Considering this apparent discrepancy, we take a different approach. Rather than modeling categories directly, we focus on the potential emergence of color categories as the result of acquiring basic visual skills. For this, we investigated whether color is represented categorically in a convolutional neural network (CNN) trained to recognize objects in natural images. We systematically trained novel output layers to the CNN for a color classification task, and found that clear borders arise between novel (non-training) colors that are largely invariant to the training colors. We confirmed these border locations by searching for the optimal border placement using an evolutionary algorithm that relies on the principle of categorical perception. Our findings also extend to stimuli with multiple, colored, words of varying color contrast, as well as colored objects with larger colored surfaces. These results provide strong evidence that color categorization can emerge with the development of object recognition.