Research

Learning does not end when behavior stops. During sleep, the brain reactivates recent experience and coordinates activity across distributed brain regions. We investigate how these offline interactions transform variable neural activity into stable and reproducible representations that support long-term memory and precise behavior.

Using long-term motor learning as a model, we study how communication among the hippocampus, prefrontal cortex, and motor cortex evolves as memories are consolidated. Early learning is characterized by broad exploration of neural population dynamics and stronger hippocampal–cortical coordination. As learning progresses, prefrontal–motor communication becomes increasingly important, while neural trajectories converge toward a stable motor manifold. Our closed-loop experiments test how specific sleep events causally shape this transition.

We also investigate how sleep reshapes neural circuits at the cellular level. In complementary studies, we examine how sleep-related hippocampal activity is linked to astrocyte-mediated synaptic remodeling in the prefrontal cortex, connecting systems-level neural dynamics with circuit plasticity during long-term memory formation.

Stable neural representations are essential for reliable behavior, but they cannot remain fixed. When behavioral demands change, the brain must reopen and reorganize previously learned neural dynamics. We investigate how consolidated representations retain the flexibility needed for adaptation.

In long-term motor learning, a new task challenge can make an established motor manifold temporarily inadequate. Neural trajectories become more variable, the system re-enters an exploratory state, and sleep-dependent hippocampal–cortical coordination is re-engaged. Through repeated learning and sleep, behavior gradually converges toward a new solution, suggesting that consolidated neural representations can be reopened and reorganized rather than simply replaced.

We also ask whether sleep allows established memories to be used prospectively. Across learning, SO–SWR coordination shifts from post-training memory reactivation toward stronger pre-training engagement of an established model. We hypothesize that this shift may reflect prospective memory simulation, allowing the brain to internally re-engage learned representations in preparation for upcoming behavior.

Together, these studies explore how the brain balances stability and flexibility—preserving useful neural structure while reopening, reorganizing, and prospectively engaging it when future behavior demands change.

If dream-related brain states have been preserved across evolution rather than eliminated, they may serve an important biological function. Yet dreaming remains one of the most difficult brain states to study objectively. We therefore seek measurable neural signatures of dream-related states across species and investigate what these internally generated dynamics may do for the brain.

A major focus of our work is PGO/P-wave–theta coupling. By comparing neural activity across rats, macaque monkeys, and humans, we investigate whether conserved interactions between PGO/P-waves and cortical theta activity can serve as a potential cross-species biomarker of dream-related sleep states.

Building on this foundation, we ask whether these internally generated brain states may help the brain recombine and reorganize existing knowledge to discover new solutions. We are developing behavioral paradigms to identify insight-like transitions and investigate how neural representations change before and after the emergence of insight and creativity.

Together, our goal is to move from identifying objective biomarkers of dream-related states toward understanding their biological function—and whether sleeping brain dynamics can contribute to the generation of new possibilities, ideas, and solutions.

We use computational and AI-based approaches to translate complex neural activity into interpretable dynamical principles. Rather than treating computation simply as a tool for data analysis, we use models to formulate mechanistic hypotheses, quantify latent neural dynamics, and generate predictions that can be tested experimentally.

Our work spans multiple levels of neural computation. We use neural manifold and population-dynamics analyses to quantify how ensemble activity changes during learning and sleep, and develop dynamical systems models to explain transitions between functional brain states. These models allow us to connect observed neural activity with underlying computational mechanisms.

We also develop AI-based and biologically interpretable latent models to reconstruct hidden neural dynamics from large-scale recordings and to study how internal representations may be reorganized during adaptation and insight. By combining theoretical models, machine learning, and causal experiments, we aim to build computational frameworks that explain not only what neural activity changes, but why those changes occur.

A stable neural representation alone is not sufficient for effective behavior. Successful movement also requires reliable transitions between distinct neural population states. We investigate how these dynamical processes are disrupted by neurological injury and how they reorganize during recovery.

Using a primate motor-recovery model, we study how movement-related neural states and their transitions recover after cortical injury. By combining longitudinal neural recordings with dynamical systems modeling, we identify computational principles that link the organization of population activity to behavioral recovery.

Ultimately, we aim to translate these principles into neurorehabilitation and brain-machine interface strategies that restore or guide functional neural dynamics after brain injury.