Priors learned across categories explain how "separable" dimensions affect human category learning

What is it about?

How people represent new categories is greatly influenced by the particular stimulus dimensions used. Categories that are aligned with separable dimensions, or those which are easily identifiable to people, are easier to learn than those that do not. Models of categorisation have explained these effects, but have not explained how these dimensions themselves are learned. We show that separable dimensions do not have to be in-built or represented explicitly: a Bayesian model that learns how stimuli are clustered across categories produces the same sort of dimensional biases that people do.

Featured Image

Photo by Jacqueline Day on Unsplash

Photo by Jacqueline Day on Unsplash

Why is it important?

REFRESH demonstrates that separable dimensions do not need to be a basic building block of cognition that category learning is built around. Instead dimensional biases fall out of learning across categories alone, making the system conceptually simpler and also responsive to the statistics of categories in the environment.