Exploring the Categorical Nature of Colour Perception: Insights from Deep Networks

This study delves into the categorical aspects of colour perception, employing the odd-one-out paradigm on artificial neural networks. We reveal a significant alignment between human data and unimodal vision networks (e.g., ImageNet object recognition). Vision-language models (e.g., CLIP text-image matching) account for the remaining unexplained data even in non-linguistic experiments. These results suggest that categorical colour perception is a language-independent representation, albeit partly shaped by linguistic colour terms during its development. Exploring the ubiquity of colour categories in Taskonomy unimodal vision networks highlights the task-dependent nature of colour categories, predominantly in semantic and 3D tasks, with a notable absence in low-level tasks. To explain this difference, we analysed kernels' responses before the winner-taking-all, observing that networks with mismatching colour categories align in continuous representations. Our findings quantify the dual influence of visual signals and linguistic factors in categorical colour perception, thereby formalising a harmonious reconciliation of the universal and relative debates.