Learning Additive and Substitutive Features.

To adapt in an ever-changing world, people infer what basic units should be used to form concepts and guide generalizations. While recent computational models of human representation learning have successfully predicted how people discover features from high-dimensional input in a number of domains (Austerweil & Griffiths, 2013), the learned features are assumed to be additive. However, this assumption is not always true in the real world. Sometimes a basic unit is substitutive (Garner, 1978), which means it can only be one value out of a set of discrete values. For example, a cat is either furry or hairless, but not both. In this paper, we explore how people form representations for substitutive features, and what computational principles guide such behavior. In a behavioral experiment, we show that not only are people capable of forming substitutive feature representations, but they also infer whether a feature should be additive or substitutive depending on the observed input. This learning behavior is predicted by our novel extension to the Austerweil and Griffiths (2011, 2013)’s feature construction framework, but not their original model. Our work contributes to the continuing effort to understand how people form representations of the world.