Using an artificial neural network to classify black-capped chickadee (Poecile atricapillus) call note types.
Résumé
The "chick-a-dee" call of the black-capped chickadee (Poecile atricapillus) contains four note types, A, B, C, and D that have important functional roles. This provides strong motivation for studying the classification of acoustic components of the call into different note types. In this paper, the spectrograms from a sample of A, B, and C notes (370 in total) were each described as a set of 9 summary features. An artificial neural network was trained to identify note type on the basis of these features, and was able to obtain better than 98% accuracy. An internal analysis of this network revealed a distributed code in which different hidden units generated high activities to different subsets of notes. By combining these different sensitivities, the network could discriminate all three types of notes. The performance of this network was compared to a discriminant analysis of the same data. This analysis also achieved a high level of performance (95%). A comparison between the two approaches revealed some striking similarities, but also some intriguing differences. These results are discussed in terms of two related issues: developing a research tool for note classification, and developing a theory of how birds themselves might classify notes.