Learning in chunks: A model of hippocampal representations for processing temporal regularities in statistical learning

We investigated the neural basis of chunking during statistical learning (SL). Behavioral evidence suggests that a common mechanism in learning and memory can serve to combine smaller units into larger ones to facilitate sensory and higher-level processing. And yet, the neural underpinnings of this mechanism remain unclear. Drawing insights from previous findings of neural codes in the hippocampus, we propose a computational model to account for the temporal chunking process in SL for sequential inputs. We operationalize chunking into a hidden Markov model (HMM) that incorporates two core principles: (1) the hidden states represent serial order rather than specific visual features, and (2) the formation of temporal chunks leads to autocorrelated brain activity. We show with numeric simulations that the HMM can decode embedded triplet representations when both assumptions hold. Applying the HMM to functional neuroimaging data from subjects performing a visual SL task, we show that decoding was successful (1) for triplet sequences but not random sequences, (2) at the later stage but not earlier stage of learning, and (3) in the hippocampus but not in the early visual cortex. These results provide evidence for a hippocampal representation of generalized temporal structure emerged from sequential visual input, shedding light on the chunking mechanism for SL. SignificanceIn statistical learning (SL), individuals develop internal representations of patterns after brief exposure to structured stimuli. People tend to recognize frequently co-occurring items as a single unit. This process, known as "chunking", is understood to play an important role in facilitating sensory processing for learning. However, its neural underpinnings remain unclear. In this study we draw insights from hippocampal coding theories and introduce a chunking model focusing on generalized presentations for SL. With functional neuroimaging data from human subjects performing a visual learning task, the chunking model successfully decoded the temporal regularities embedded in the sequential inputs. This model and related findings provide critical evidence for a chunking process underlying SL as well as its representation in the human hippocampus.