Hierarchical Color Learning in Convolutional Neural Networks

Chris Hickey, Byoung-Tak Zhang; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 384-385

Empirical evidence suggests that color categories emerge in a universal, recurrent, hierarchical pattern across different cultures. This pattern is referred to as the "Color Hierarchy". Over two experiments, the present study examines whether there is evidence for such hierarchical color category learning patterns in Convolutional Neural Networks (CNNs). Experiment A investigated whether color categories are learned randomly, or in a fixed, hierarchical fashion. Results show that colors higher up the Color Hierarchy (e.g. red) were generally learned before colors lower down the hierarchy (e.g. brown, orange, gray). Experiment B examined whether object color affects recall in object detection. Similar to Experiment A, results found that object recall was noticeably impacted by color, with colors higher up the Color Hierarchy generally showing better recall. Additionally, objects whose color can be described by adjectives that emphasise colorfulness (e.g. Vivid, Brilliant, Deep) show better recall than those which de-emphasise colorfulness (e.g. Dark, Pale, Light). These results highlight similarities between humans and CNNs in color perception, and provide insight into factors that influence object detection.