New VAAE Journal Publication

A new article from members of the VAAE team has been published in The Journal of the Acoustical Society of America (Vol.149, Issue 6).

The paper "Matrix analysis for fast learning of neural networks with application to the classification of acoustic spectra" by Vlad Paul and Prof. Philip Nelson describes a novel technique to reduce the training time of neural networks based on the Singular Value Decomposition of the weight matrices.

 

Neural networks are increasingly being applied to problems in acoustics and audio signal processing. The performance of machine learning algorithms strongly depends on the size of the dataset and using large datasets require a large amount of training time. This paper shows that the weights of neural networks can be approximated by the Singular Value Decomposition and the dimensions of the resulting matrices can be reduced in a consecutive way during training, by eliminating small singular values in a meaningful way. The simulation results show that the technique can reduce the training time by 1/3, without loosing any performance. Therefore, using this approach a reduction in training time can be achieved whilst retaining the accuracy of the generated weights.