Machine Learning for Bird Song Learning

Bird song is a complex, multidimensional trait, varying in pitch, temporal features and sometimes in spectral complexity too. If we want to understand what variation in songs means to birds, we have to first understand how such variation is perceived. In this project, we used an operant conditioning protocol to assess birds' judgments of song differences. We developed new operant conditioning equipment ("Operanter") based on Raspberry Pi hardware, and using PIT tags worn on birds' legs. This allowed us to carry out operant conditioning in aviary groups of birds, and with free-living wild birds - and collect large numbers of birds' judgements.

We tied this to a machine-learning approach to song syllable comparison: we trained machine learning algorithms to differentiate between song syllables in the same way that the birds themselves did.

And finally, we are in the process of applying song comparison algorithms, "tuned" by avian perception, to examine processes of cultural evolution across species.