Adaptive feedback control as a mechanism for sequence learning

Humans and other animals can rapidly learn sequences of events and subsequently retrieve them with minimal repetition. By recording neuronal activity from four brain areas in human epilepsy patients performing a sequence learning task, we investigated how neuronal populations support this process. We examined two candidate sequencing algorithms, inspired by computer science: indexing and queuing. During learning, population activity occupied subspaces consistent with queuing and credit assignment. Once the sequence was fully learned, however, the overall neural geometry reorganized: the queue subspace became nonlinear, the enqueue subspace diminished, and the representation shifted toward an indexing-like format. Using network modeling, we further explored the governing principles underlying this transition, demonstrating that adaptive feedback control provides a plausible mechanism for the modulation of neural geometry during learning. Finally, we found that, upon successful learning, the hippocampus broadcasts sequence information to other regions through a low-dimensional subspace. Together, these findings suggest that sequence learning can be understood in terms of algorithmic neural subspaces and their dynamic reorganization across the course of learning.