Generative neural network autoencoder for novel song production and visualization of passerine birdsong

Students

College

Faculty

Associate Professor
College

Project Summary

Passerine songbirds are integral to fields of neuroscience. The specialized neural circuit, the song system, is the source of active signalling from males but is present in both males and females, albeit females have been viewed as signal receivers and choosers. This implies that the song system has utility not confined to that of courtship signalling.
 
Male songbirds sing in varying potency, eliciting different rates of CSD response. This trait generates selectivity in female songbirds, and the precise neural circuitry for the behaviour remains poorly understood. That is to say, the effect of male birdsong subtleties on female courtship behaviour are unknown to date, as we do not know specifically which song qualities the females use in order to make their mating choices.
 
Manipulation is typically executed either manually through guessing parameters or through creating an autoencoder for creating a representation of a high potency song and low potency song, allowing for the generation of intermediate potency of song based on custom parameters. As such, this project strives to parameterize birdsong in order to allow manipulation of male song features to determine female preference. In this way, the type of features could be effectively controlled. 
 
In order to classify and thereby represent the patterns in passerine birdsong, this project strives to visualize sequences of inputted birdsong through two generative autoencoder networks (multidimensional scaling and visualization), of which is then clustered into categories via an unsupervised UMAP/HBDSCAN algorithm. Clustering of categories, which is an unsupervised learning method in which the data is sorted into similar groups, ultimately allows further understanding of birdsong corpus as we can match song patterns with associated bird behaviour. 
 
Through my summer research experience, I was able to learn how to fundamentally design and execute an idea. I gained profuse exposure to applications of machine learning, mathematics, statistics, and computer science to biological problems—an unconventional but extremely useful approach to interdisciplinary research between computer science and biology.