A mixture of generative models strategy helps humans generalize across tasks

Abstract

What role do generative models play in generalization of learning in humans? Our novel multi-task prediction paradigm--where participants complete four sequence learning tasks, each being a different instance of a common generative family--allows the separate study of within-task learning (i.e., finding the solution to each of the tasks), and across-task learning (i.e., learning a task differently because of past experiences). The very first responses participants make in each task are not yet affected by within-task learning and thus reflect their priors. Our results show that these priors change across successive tasks, increasingly resembling the underlying generative family. We conceptualize multi-task learning as arising from a mixture-of-generative-models learning strategy, whereby participants simultaneously entertain multiple candidate models which compete against each other to explain the experienced sequences. This framework predicts specific error patterns, as well as a gating mechanism for learning, both of which are observed in the data.