The implications of null patterns and output unit activation functions on simulation studies of learning: A case study of patterning

Feature weighting in "chick-a-dee" call notes of Poecile atricapillus

Artificial neural networks were trained to discriminate between different note types from the black-capped chickadee (Poecile atricapillus) "chick-a-dee" call. Each individual note was represented as a vector of summary features taken from note spectrograms and networks were trained to respond to exemplar notes of one type and to fail to respond to exemplar notes of another type. Following initial network training, the network was presented novel notes in which individual acoustic features had been modified. The strength of the response of the network to each novel and shifted note was recorded. When network responses were plotted as a function of the degree of acoustic feature modification and training context, it became clear that modifications of some acoustic features had significant effects on network responses, while others did not. Moreover, the training context of the network also played a role in the responses of networks to manipulated test notes. The implications of using artificial neural networks to generate testable hypotheses for animal research and the role of context are discussed.