Daniel Levenstein, PhD

Like learning, sleep changes the brain to improve its future performance. Unlike learning, these changes occur in the absence of overt behavior or sensory input. This “offline learning” thus contains a mystery: how does spontaneous neural activity, which is entirely self-organized, improve brain function? My lab aims to solve this mystery, and develop theories of offline learning that can be used to mimic its computational benefits in artificial neural networks and understand its disruption in neuropsychiatric disorders.

The sleeping brain contains a zoo of complex activity patterns, commonly called sleep oscillations, with rich representational content and emergent spatiotemporal dynamics that coordinate neuronal interactions within and between brain regions. One well-studied example is sharp wave-ripples (SWRs): high-frequency oscillations in the hippocampus during which neural activity simulates wake-like trajectories that “replay” previous experience. SWRs interact with sleep oscillations across the brain, most notably slow oscillations (SW) in the neocortex. Critically, SWR-SW interactions support the consolidation of recent experiences into long-term memory, and are disrupted in a number of neuropsychiatric disorders including schizophrenia, Alzheimer's disease, depression, epilepsy, and autism spectrum disorder.

The work in my lab centers around three questions, using hippocampal SWRs and replay as a case study for offline learning: “How does spontaneous activity emerge and self-organize in neural networks?”, “How does plasticity during spontaneous activity change the brain?”, and “How do those changes improve the brain’s operations and performance on future tasks?”. To answer these questions, we use artificial neural network (ANN) models, dynamical systems theory, and neural data analysis – working closely with experimental collaborators to inspire the design of our models and to ground them in experimental data. This NeuroAI approach, in which brain-inspired ANNs are built and used as models for the brain, is particularly well-suited to bridge neurons’ circuit and cellular-level properties with their cognitive and behavioral implications.