Our environment contains temporal information unfolding simultaneously at multiple timescales. How do we learn and represent these dynamic and overlapping information streams? We investigated these processes in a statistical learning paradigm with simultaneous short and long timescale contingencies. Human participants ( n = 96) played a game where they learned to quickly click on a target image when it appeared in one of nine locations, in eight different contexts. Across contexts, we manipulated the order of target locations: at a short timescale, the order of pairs of sequential locations in which the target appeared; at a longer timescale, the set of locations that appeared in the first versus the second half of the game. Participants periodically predicted the upcoming target location, and later performed similarity judgments comparing the games based on their order properties. Participants showed context-dependent sensitivity to order information at both short and long timescales, with evidence of stronger learning for short timescales. We modeled the learning paradigm using a gated recurrent network trained to make immediate predictions, which demonstrated multilevel learning timecourses and patterns of sensitivity to the similarity structure of the games that mirrored human participants. The model grouped games with matching rule structure and dissociated games based on low-level order information more so than high-level order information. The work shows how humans and models can rapidly and concurrently acquire order information at different timescales.

Our understanding of the world is shaped by inferences about underlying structure. For example, at the gym, you might notice that the same people tend to arrive around the same time and infer that they are friends that work out together. Consistent with this idea, after participants are presented with a temporal sequence of objects that follows an underlying community structure, they are biased to infer that objects from the same community share the same properties. Here, we used fMRI to measure neural representations of objects after temporal community structure learning and examine how these representations support inference about object relationships. We found that community structure learning affected inferred object similarity: When asked to spatially group items based on their experience, participants tended to group together objects from the same community. Neural representations in perirhinal cortex predicted individual differences in object grouping, suggesting that high-level object representations are affected by temporal community learning. Furthermore, participants were biased to infer that objects from the same community would share the same properties. Using computational modeling of temporal learning and inference decisions, we found that inductive reasoning is influenced by both detailed knowledge of temporal statistics and abstract knowledge of the temporal communities. The fidelity of temporal community representations in hippocampus and precuneus predicted the degree to which temporal community membership biased reasoning decisions. Our results suggest that temporal knowledge is represented at multiple levels of abstraction, and that perirhinal cortex, hippocampus, and precuneus may support inference based on this knowledge.

Despite the importance of learning and remembering across the lifespan, little is known about how the episodic memory system develops to support the extraction of associative structure from the environment. Here, we relate individual differences in volumes along the hippocampal long axis to performance on statistical learning and associative inference tasks—both of which require encoding associations that span multiple episodes—in a developmental sample ranging from ages 6 to 30 years. Relating age to volume, we found dissociable patterns across the hippocampal long axis, with opposite nonlinear volume changes in the head and body. These structural differences were paralleled by performance gains across the age range on both tasks, suggesting improvements in the cross-episode binding ability from childhood to adulthood. Controlling for age, we also found that smaller hippocampal heads were associated with superior behavioral performance on both tasks, consistent with this region's hypothesized role in forming generalized codes spanning events. Collectively, these results highlight the importance of examining hippocampal development as a function of position along the hippocampal axis and suggest that the hippocampal head is particularly important in encoding associative structure across development.

The sensory input that we experience is highly patterned, and we are experts at detecting these regularities. Although the extraction of such regularities, or statistical learning (SL), is typically viewed as a cortical process, recent studies have implicated the medial temporal lobe (MTL), including the hippocampus. These studies have employed fMRI, leaving open the possibility that the MTL is involved but not necessary for SL. Here, we examined this issue in a case study of LSJ, a patient with complete bilateral hippocampal loss and broader MTL damage. In Experiments 1 and 2, LSJ and matched control participants were passively exposed to a continuous sequence of shapes, syllables, scenes, or tones containing temporal regularities in the co-occurrence of items. In a subsequent test phase, the control groups exhibited reliable SL in all conditions, successfully discriminating regularities from recombinations of the same items into novel foil sequences. LSJ, however, exhibited no SL, failing to discriminate regularities from foils. Experiment 3 ruled out more general explanations for this failure, such as inattention during exposure or difficulty following test instructions, by showing that LSJ could discriminate which individual items had been exposed. These findings provide converging support for the importance of the MTL in extracting temporal regularities.

Human and animal lesion studies have shown that behavior can be catastrophically impaired after bilateral lesions but that unilateral damage often produces little or no effect, even controlling for lesion extent. This pattern is found across many different sensory, motor, and memory domains. Despite these findings, there has been no systematic, computational explanation. We found that the same striking difference between unilateral and bilateral damage emerged in a distributed, recurrent attractor neural network. The difference persists in simple feedforward networks, where it can be understood in explicit quantitative terms. In essence, damage both distorts and reduces the magnitude of relevant activity in each hemisphere. Unilateral damage reduces the relative magnitude of the contribution to performance of the damaged side, allowing the intact side to dominate performance. In contrast, balanced bilateral damage distorts representations on both sides, which contribute equally, resulting in degraded performance. The model's ability to account for relevant patient data suggests that mechanisms similar to those in the model may operate in the brain.