Computations in the Ascending Auditory Pathway of Songbirds Related to Song Learning
In general, the song of each adult songbird is unique to that individual. Although some of the parameters of song result from maturational processes which are innately specified, the spectral and temporal parameters of song that confer individual specificity are acquired by learning. The specifics of an individual's song are reflected in idiosyncratic physiological properties of auditory neurons in song system nuclei such as the nucleus hyperstriatum ventrale pars caudale (HVc). The authors are exploring the transformation from classical auditory to song-specific neuronal properties. This transformation is a complex one which to date has not yielded to classical neuronanatomical and neurophysiological methods of analysis. Furthermore, the idiosyncratic nature of song system auditory response properties makes it difficult to collapse data across individuals. To address these issues, the authors have investigated a connectionist modelling approach. Their initial efforts have been directed at auditory neurons of the thalamus. In the thalamus, simple linear-static models that use average firing rates of responses to tone bursts have proven to be poor predictors of responses to song. In contrast, time-delay neural network (TDNN) architectures that are trained with time-varying responses to tone bursts using the backpropagation algorithm make excellent predictions of the responses to songs. The authors are currently extending these architectures in an attempt to capture salient features of the responses of song system auditory neurons.