Building a state space for song learning

摘要: The songbird system has shed light on how the brain produces precisely timed behavioral sequences, and implements reinforcement learning (RL). RL is a powerful strategy for what action to produce in each state, but requires unique representation of states involved task. Songbird circuitry thought operate using moment within song syllables, consistent with sparse sequential bursting neurons premotor cortical nucleus HVC. However, such sequences are not present very young birds, which sing highly variable syllables random lengths. Here, we review expand upon model could construct latent support RL, new data elucidating connections between HVC auditory areas. We hypothesize that occurs via four distinct plasticity processes: 1) formation ‘tutor memory’ areas; 2) appropriately-timed seeded by inputs from areas spontaneously replaying tutor song; 3) strengthening, during spontaneous replay, corresponding timing sequence, aligning motor representations subsequent evaluation; 4) strengthening output desired sounds, well-described circuitry.