Reinforcement Learning and Birdsong

The production of a birdsong is reminiscent of a central pattern generator, at least for birds with a singing repertoire: a neural circuit produces a (relatively) stereotypic motor output (McCasland, 1987). Nevertheless, there are some important differences: the circuitry involved in birdsong is distributed across many areas of the brain; some species (such as zebra finch) learn their songs under natural conditions—without learning, song output is degraded. The analysis of such learning touches on some ideas important in artificial intelligence, specifically reinforcement learning (“RL”), ideas also advanced to explain some forms of mammalian learning (rewarded and unrewarded behaviors). Birdsong is furthermore important for insights it might provide into human language production. Birdsong and speech have underlying rhythms (Gehrig et al., 2019), reminiscent of CPGs, but of course for song and speech, there must be selective structure imposed on the underlying rhythm. How that structure comes to have its own special properties raises problems that are independent of how the rhythm is generated.