Building integrated representations through interleaved learning.

Abstract

Inferring relationships that go beyond our direct experience is essential for understanding our environment. This capacity requires either building representations that directly reflect structure across experiences as we encounter them or deriving the indirect relationships across experiences as the need arises. Building structure directly into overlapping representations allows for powerful learning and generalization in neural network models, but building these so-called distributed representations requires inputs to be encountered in interleaved order. We test whether interleaving similarly facilitates the formation of representations that directly integrate related experiences in humans and what advantages such integration may confer for behavior. In a series of behavioral experiments, we present evidence that interleaved learning indeed promotes the formation of representations that directly link across related experiences. As in neural network models, interleaved learning gives rise to fast and automatic recognition of item relatedness, affords efficient generalization, and is especially critical for inference when learning requires statistical integration of noisy information over time. We use the data to adjudicate between several existing computational models of human memory and inference. The results demonstrate the power of interleaved learning and implicate the formation of integrated, distributed representations that support generalization in humans. Public Significance StatementThe study provides evidence that presenting information in an intermixed order facilitates the storage of this information in a format that promotes the efficient discovery of hidden relationships.

An inductive bias for slowly changing features in human reinforcement learning

“…This raises the possibility that slowness, not only in feature dynamics but also in task rules, may aid learning. However, it is worth noting that interleaved training might promote the formation of more generaliseable representations [63], suggesting that the optimal learning curriculum may differ depending on the task at hand. In sum, multiple lines of research point toward a beneficial effect of slowness on learning.…”

A neural network model of hippocampal contributions to category learning

“…This suggests that a control mechanism that enhances one pathway over another depending on the task would be beneficial for behavior. In a recent paper, we adopted a version of C-HORSE that invoked such a control function in order to explain behavior across tasks with different demands in an associative inference paradigm ( Zhou et al, 2023 ). Medial prefrontal cortex could potentially carry out a control function of this kind ( Sherman et al, 2023 ), as it participates in category learning ( Mack et al, 2020 ) and is known to modulate CA1 representations as a function of task ( Eichenbaum, 2017 ; Guise and Shapiro, 2017 ).…”

“…We adopted a neural network model of the hippocampus developed after a lineage of models used to explain how the DG, CA3, and CA1 subfields of the hippocampus contribute to episodic memory ( Ketz et al, 2013 ; Norman and O’Reilly, 2003 ; O’Reilly and Rudy, 2001 ). This variant, C-HORSE, was developed recently to account for the role of the hippocampus in statistical learning ( Schapiro et al, 2017b ; Zhou et al, 2023 ). Simulations were performed in the Emergent simulation environment (version 7.0.1, Aisa et al, 2008 , O’Reilly et al, 2014a ).…”

“…In the present work, we ask what computational properties of the hippocampus might allow it to contribute to category learning. Using a neural network model of the hippocampus named C-HORSE (Complementary Hippocampal Operations for Representing Statistics and Episodes), we previously demonstrated how the hippocampus might contribute to learning temporal regularities embedded in continuous sequences of stimuli (temporal statistical learning) and to inference over pairwise associations ( Schapiro et al, 2017b ; Zhou et al, 2023 ). We showed that the heterogeneous properties of the two main pathways within the hippocampus may support complementary learning systems, with one pathway specializing in the rapid encoding of individual episodes and another in extracting statistics over time.…”

Memory consolidation during rest forms shortcuts in a cognitive map

“…Given the proposed role of replay in reorganising memory to extend the cognitive map beyond direct experience, and evidence to suggest that TMR can bias the content of replay, we predicted that TMR might provide a tool for investigating the consequences of reorganising memory during periods of rest. Thus, we sought to test the hypotheses that TMR improves inferential choice and that this improvement is driven by the formation of new inferred links (or 'shortcuts') between cues that have not been experienced together 22,30 .…”