Development and validation of a computational model of higher-order statistical learning on graphs in humans

点击翻译按钮获取中文摘要

As humans navigate their environment, anticipation, planning, and perception all require an accurate map of the statistical regularities governing their visual, linguistic, auditory, and social experiences. In each context, hu- man experience consists of a sequence of events. Each event succeeds another according to a set of underlying rules codifying possible event-to-event transitions, and the likelihood of each. To make predictions about the fu- ture and respond to the environment with flexible behavior, humans must infer this network of transitions, forming a cognitive map of causes and effects. Such maps and inferences are made possible by statistical learning. The study of statistical learning represents a major opportunity for computational psychiatry for three reasons. First, statistical learning shows differential accuracy across psychiatric conditions, task domains, and temporal scales of experience. Second, statistical learning has marked potential for back-translation; multiple features of statistical learning behavior and its neural underpinnings are conserved in non-human primates, and simpler forms of sequence learning exist in other mammals (rats and mice) as well as birds. Third, – as we describe in depth in our proposal – statistical learning can be formally modeled mathematically. It is now timely to develop a flexible computational model of statistical learning. To serve the goals of com- putational psychiatry, the functional form of such a model should reflect general principles of statistical learning and the parameters should be sensitive to variability in behavior across the many specific disorders where deficits appear. In preliminary experimental, computational, and theoretical work, we have uncovered a novel behavioral signature of statistical learning; we have also translated that behavior into a formal model – inspired by principles of statistical physics – with mathematically well-defined parameters, thereby deriving a theory that is grounded in our previous experimental findings. Finally, we have experimentally validated the model by making accurate predictions of behavior in a novel experiment. Here we assemble a complementary set of co-investigators who have co-authored 31 papers in pairs or triplets, with expertise in mathematical modeling and statistical physics (Bassett), statistical models of behavior (Moore), intensive longitudinal experiments (Lydon-Staley), statistical learning (Thompson-Schill), and sensory process- ing in psychiatry (Wolf). Together, we offer a well-integrated theoretical and experimental plan to hone our math- ematical model of an aspect of human behavior that has not been extensively analyzed computationally, and in which the underlying dimensional process is affected in psychiatric disorders. We distill our aims into reliability, relevance, and generalizability of our model. Our approach is three-pronged, with innovations in experiment, computation, and theory building on our team’s diverse expertise. Each prong will address all three aims, thereby integrating our efforts to build a computational model of statistical learning behavior supporting future advances in computational psychiatry. Our proposed efforts provide the foundation for an R01 extending to patients.