High-level Musical Feature Learning using Neural Networks
Keunwoo Choi, Centre for Digital Music, Queen Mary University of London
High-level musical feature such as mood tags can be learned using deep convolutional neural networks. Unlike conventional approaches, features can be learned automatically by minimising cost function, which can be defined according to evaluation criteria.
The demonstrator presents an analysis of the learned features by a technique named deconvolution, which enables us to see and listen the learned features. Two convolutional neural networks – one for genre prediction and the other for automatic tagging – will be analysed.
The learned networks will be used as a feature extractor in playlist generation algorithm for content-based music recommendation system.