Replay-driven task orthogonalization and abstraction for continual learning

The National Institute of Mental Health awarded New York University $226,404 on September 3, 2026, under the Mental Health Research Grants program (CFDA 93.242) to conduct experimental research examining how memory replay supports continual learning in the brain.

The research tests two hypotheses: that the brain forms representations segregating context- and task-specific information while abstracting shared structure across tasks, and that replay supports continual learning by orthogonalizing and abstracting task representations. Participants will learn one-step transitions before planning longer action sequences across three different graphs while brain activity is recorded using magnetoencephalography (MEG). The study analyzes both behavioral and neural patterns to understand how the brain balances keeping task-specific details distinct while integrating shared structure to enable generalization.

The award carries a period of performance through March 1, 2029, with place of performance in New York, New York.

Abstract

One of the brain's most remarkable abilities is its capacity to learn and adapt continuously throughout life. This capacity relies on balancing two competing demands: keeping context- and task-specific details distinct while also integrating shared structure across experiences to enable generalization. However, we still do not fully understand how the brain concurrently manages these demands, largely due to the lack of tasks designed for studying them together. We propose that memory replay-the reactivation of brain activity patterns in the absence of overt task demand-support continual learning in the brain by reorganizing neural representations to fulfill these demands. Although past research has identified evidence of replay across animals and humans, its role in learning and behavior remain unclear. To address these gaps, we have developed a new experimental design that examines examines how the brain manages to keep task-specific details separate while also extracting common patterns during continual learning. We will also explore how memory replay enables the brain to strike this balance. Our approach tests two main hypotheses: first, that the brain forms representations that segregate context- and task- specific information while abstracting shared structure across tasks; and second, that replay supports continual learning by helping to orthogonalize and abstract task representations. In our study, participants will first learn simple, one-step transitions before planning longer action sequences across three different graphs, all while their brain activity is recorded using magnetoencephalography (MEG). This design will allow US to assess how well participants retain details specific to the individual graphs and extract a hidden, abstract structure common to all of them. We will analyze both behavior and neural patterns to understand how the brain manages these dual demands, and we will compare human behavior and neural representations with that of neural network models optimized for the same tasks. We will also use advanced MEG decoding techniques to track replay events during both rest and active task phases, examining how these events shape behavior and task representations. Complementary measures, such as eye-tracking, will help US explore how different physiological states influence replay dynamics. By combining behavioral testing, neuroimaging, and computational modeling, this study aims to provide new insights into how the brain continually adapts to changing environments. The findings will deepen our fundamental understanding of human learning and memory, and guide future efforts to enhance cognitive function in educational, clinical, and aging settings.