Distributed representation of discrete sequential vocalization in the Bengalese finch (Lonchura striata var. domestica)

Bengalese finches learn and produce sequential vocalizations with a complex song syntax. Various models for the song syntax have been proposed, each of which focuses on several different characteristics of the syntax. However, methods to model these multiple characteristics in a single framework have not been well studied. Here, we propose a model that explains three prominent characteristics of the song syntax in Bengalese finches in a single unified framework. First, the generation of a vocal element depends on multiple preceding elements. Second, a song often contains repetitions of a single element type. Third, a song often begins with a special sequence called an introductory sequence. In this study, an effective way was sought to model these three characteristics in the framework of a conditional probability in symbol sequences. The model takes a distributed representation of a preceding sequence as an input, which is defined by a set of decaying values activated when the vocal elements are generated. The proposed model is shown to outperform conventional syntax models in predicting sequences in novel songs. The results suggest that the song syntax of the bird’s brain might also be represented by decaying activities of populations of neurons.