Efficient design of real- and complex-valued neural networks

Abstract

This work investigates the application of neural networks to problems in acoustics. the aim of the work is to develop a strategy for designing effective neural networks whilst ensuring their computational efficiency. The approach taken is to use the singular value decomposition (SVD) of the weight matrices in the network to develop a strategy for pruning the network. The SVD - based approach iteratively removes neurons from the hidden layer during the training such that the final trained model consists of a hidden layer with reduced dimensions. The pruning technique is applied to real- and complex valued networks focusing on the multilayer percetrons (MLPs) and the current neural networks (RNNS). The theory underpinning each real- and complex-valued network model is discussed comprehensively and their performance is investigated for a number of audio signal processing tasks. Results show that the SVD-based pruning technique can be successfully applied to both real- and complex-valued network models regardless of the model architecture for both classification and regression tasks. A direct comparison between real- and complex-valued network architecture shows the potential of complex-valued machine learning models and their potential benefits for audio processing.

More information

Identifiers

ORCID for Vlad Stefan Paul: orcid.org/0000-0002-5562-6102

ORCID for Philip Nelson: orcid.org/0000-0002-9563-3235

ORCID for Filippo Fazi: orcid.org/0000-0003-4129-1433

Catalogue record

Export record