Abstract
Experience is known to facilitate our ability to extract regularities from simple repetitive patterns to more complex probabilistic combinations (e.g. as in language, music, navigation). However, little is known about the neural mechanisms that mediate our ability to learn hierarchical structures. Here we combine behavioral and functional MRI measurements to investigate the human brain circuits involved in learning of hierarchical structures. In particular, we employed variable memory length Markov models to design hierarchically structured temporal sequences of increasing complexity. We first trained observers with sequences of four symbols that were determined by their probability of occurrence (0.2, 0.8, 0, 0) and then sequences determined by their temporal context. We measured performance and fMRI responses before and after training on these two sequence types. In each trial, we presented observers with a sequence of symbols followed by a test stimulus. Observers were asked to indicate whether the test stimulus was expected or not. Our results demonstrate that dissociable brain circuits are involved in learning regularities determined by frequency of occurrence vs. temporal context. In particular, learning occurrence probabilities engaged frontal (inferior and middle frontal gyrus), parietal (inferior parietal lobule) and superior temporal regions, while learning temporal context engaged frontal (superior and medial frontal gyrus, cingulate) and subcortical circuits (putamen). Fronto-parietal regions showed increased fMRI responses to structured compared to random sequences early in training while decreased responses after training. In contrast, in subcortical regions, higher responses were observed for structured compared to random sequences only after training. Our results are consistent with the role of fronto-parietal circuits in identifying novel patterns, and the involvement of subcortical regions in contextual learning. Thus, our findings suggest that learning hierarchical structures is implemented by fast learning of frequency statistics in fronto-parietal regions, while conditional probability learning in subcortical regions later in training.
Meeting abstract presented at VSS 2015