Conditional Restricted Boltzmann Machines for Mono/Polyphonic Composer Identification

Research output: Unpublished contribution to conferenceUnpublished paper

93 Downloads (Pure)


In this paper, the effectiveness of Conditional Restricted Boltzmann Machines (CRBMs) as universal feature extractors for classifying symbolic music is investigated. An average monophonic classification accuracy of 72% was achieved when discriminating between string quartets movements of Mozart and Haydn. When the decisions of individual monophonic parts were combined using a basic voting scheme, a polyphonic classification accuracy of 96% was achieved, a substantial improvement compared to the best state-of-the-art performance to date of 80%. It was observed that the classification rate depended heavily on the exact composition of training and test set: the classifier performance deteriorated when the time of composition increased between training and test set. This supports the observation that the ”style” of a composer is not a fixed given but varies over time.
Original languageEnglish
Publication statusPublished - 6 Nov 2015
Event27th Benelux Conference on Artificial Intelligence - BNAIC 2015 - University of Hasselt, Hasselt, Belgium
Duration: 5 Nov 20156 Nov 2015


Conference27th Benelux Conference on Artificial Intelligence - BNAIC 2015

Cite this