UC Irvine

In order to organize the information that people encounter in their environments they have to develop the ability to organize their experience into categories. Category learning has been studied using both stable environments in which the underlying categories do not change, and in dynamic environments in which the underlying category structure changes independent of people’s behavior. In many natural environments, however, category structures can change as a consequence of people’s behavior. In order to study how people learn categories that are dynamically coupled to their behavior, we conduct two experiments in which the underlying category structure changed as people became more accurate. Results from these experiments show that people can quickly adapt to changes in an underlying category structure, and that complicated learning dynamics occur as they transition from one structure to another.

We argue that existing models of category learning are not well suited to capturing people’s behavior in these tasks, and introduce a new framework to measure the dynamics of categorization. This approach is based on Coupled Hidden Markov Models (CHMMs), and allows us to directly model the effect of stimulus similarity in the categorical associations of stimuli across trials in a task. Using data from a previously published category learning study, we show that the CHMM can adequately describe people’s categorization behavior, and serves as an interpretable and useful measurement model for understanding how people learn to associate stimuli with categories over time.

We further test the CHMM as a predictive model with out-of-sample generalization tests, using data from two previous experiments. One experiment involves people learning to categorize faces in terms of different category structures including gender, hair color, and trustworthiness. We show that the CHMM performs progressively better for more abstract categories, and often outperforms the well-established Generalized Context Model. The second experiment involves simple perceptual stimuli learning in a training-transfer design. We show that the CHMM again matches or outperforms the Generalized Context Model in predictive accuracy for transfer stimuli. We argue that the key advantage of the CHMM is the flexibility that motivated its development as a measurement model. While the GCM assumes people always know the true category structure, the CHMM has the flexibility to infer that some people at some stages of learning have incorrect category structures. We conclude with a discussion of future avenues for experimentation and model development that will improve our understanding of how people learn to make categorization decisions in a changing environment.