Hippocampal Generative Processes During Sleep Support Rapid Learning of New Associations

Internally generated representations (of things like events or facts not currently present) were framed by George Dragoi, MD, PhD, as a shared substrate for memory, planning, imagining, dreaming, and inference, with the hippocampus critically involved in both humans and rodents. When abnormal, these representations may take the form of hallucinations and delusions in disorders like schizophrenia. Dragoi shared these insights at the Precision Therapeutics in Schizophrenia symposium at the Fralin Biomedical Research Institute.

Neuronal activity is organized like a generative grammar, Dragoi proposed. Beyond single-neuron activity, short sequential motifs termed tuplets—2 to 4 neurons that fired in the same order much more often than chance—served as modules, analogous to words, for building longer sequences.¹ According to Dragoi, "there are extended sequences that are made to represent new information, new experiences, but they are based on these building blocks which are then later modified by synaptic plasticity to then be consolidated and form memories of events experienced."

The role of this process in insight learning has been tested in rats.² The paradigm modeled problem solving after sleep: animals learned 6 abstract flavor-location associations over weeks, forming a mental schema, and then received electrodes over the hippocampus and the medial prefrontal cortex. When the same rule applied to new flavor cues at the same locations, performance was poor during initial exposure but improved right away after 1 hour of offline sleep or rest—far faster than the original learning. Recordings showed that new associations between the new cues and the old locations formed during those offline sessions.

Electrical stimulation that blocked these hippocampal epochs impaired rapid learning across multiple sessions, despite repeated exposure to the new cues. Rats left to sleep overnight still solved the task the next day, indicating that slower learning remained intact. Dragoi described this as one of the first demonstrations that generative processes were important for rapid learning. He emphasized the importance of sufficient sleep, describing it as an active process for the brain. "This is an active brain state where a lot of new associations and reconfigurations in the brain network that support learning occur," he said.

Dr Dragoi is a professor of psychiatry and neuroscience at Yale University School of Medicine.

References

1. Liu K, Sibille J, Dragoi G. Generative predictive codes by multiplexed hippocampal neuronal tuplets. Neuron. 2018;99(6):1329-1341.

2. Bhattarai B, Dragoi G. Offline generative network reconfiguration guides insight-like accelerated learning by assimilation into schema in rats. Nat Commun. 2026;17(1):9694.