Deep Gaussian Processes for Estimating Music Mood

Abstract

Due to the subjective nature of music mood, it is challenging to computationally model the affective content of music. In this work for estimating music mood, we present a deep Gaussian process regression model that is inspired from deep learning architectures. A variational Bayesian approach is used to learn the Gaussian mixture model representation for acoustic features. Since the exact inference on deep Gaussian processes (DGP) is intractable, the pseudo-data approximation is used to reduce the training complexity and Monte Carlo sampling technique is used to solve the intractability problem during training. A detailed derivation of a 3-layer DGP is presented that can be easily generalized to a L-layer DGP. The proposed work is evaluated on PMEmo dataset containing valence and arousal annotations of Western popular music and achieves an improvement in the coefficient of determination of 16.9% for arousal and 28.8% for valence estimation relative to the baseline single-layer Gaussian process.