Investigating Nonnegative Autoencoders for Efficient Audio Decomposition

Nonnegative Matrix Factorization (NMF) is a powerful technique for decomposing a music recording's magnitude spectrogram into musically meaningful spectral and activation patterns. In recent years, musically informed NMF-based audio decomposition has been simulated using neural networks, which opens up new paths of exploiting recent deep learning frame-works, including libraries for efficient gradient computations. In this article, we continue this strand of research by considering Nonnegative Autoencoders (NAE) in combination with gradient projection and structured dropout techniques. Conducting experiments based on piano recordings, we compare the decomposition results of NAE-based approaches with those obtained from a score-informed NMF variant. In this context, we examine various gradient descent methods using fixed and adaptive learning rates for deriving the NAE encoder and decoder parameters. Among others, we show how the famous multiplicative update rules for NMF can be transferred to the case of NAEs. The overall goal of our contribution is to illustrate the benefits and limitations of the various techniques concerning implementation issues (CPU, GPU), convergence speed, and overall runtime.