Basal ganglia-dependent expression of recent song learning in the juvenile finch

Many of our most impressive skills, from speech to music, evolve across many dimensions and are learned spontaneously without external reinforcement. While we know much about how the basal ganglia contribute to learning low-dimensional, externally reinforced behaviors like lever pressing, their contribution to learning high-dimensional behaviors is less clear. Birdsong learning is a paradigmatic example of high-dimensional learning. Juvenile songbirds vary many acoustic dimensions as they learn to copy their tutor’s song, a spontaneous process that evolves without external reinforcement. While songbirds possess a song-specialized basal ganglia nucleus (the sBG) necessary for song copying, nearly all of what we know about the sBG derives from studies in adult birds that examine a unidimensional, externally reinforced type of “vocal lever pressing.” We combined sophisticated computational analyses with closed-loop optogenetic suppression to directly test the hypothesis that sBG activity is acutely necessary to express recently learned changes during juvenile song copying.

The high dimensionality of birdsong is one major challenge to testing this hypothesis. To address this challenge, we applied a variational autoencoder to compress high-dimensional vocal data into a low-dimensional latent space, then developed a feed forward neural network that predicted the age of production for every syllable rendition based on these latent dimensions. In this way, we identified learning-relevant dimensions in the vocal data. We combined this with closed-loop optogenetic suppression of sBG activity in juvenile zebra finches as they copied their tutor’s song. sBG suppression induced a regression in song learning immediately and transiently (i.e., only on targeted renditions), erasing on average the previous 8 hours of learning. This indicates that, during juvenile song copying, the sBG functions to express recently acquired learning on a rendition-to-rendition basis. Our work is of broad relevance to understanding how the basal ganglia help solve high-dimensional learning problems, even without external reinforcement.