Effects of Spectral Radius on Echo-State-Network's Training

Abstract

The echo-state-network approach for training recurrent neural networks can yield good results. However, the results depend on the experience of neural network design. It usually requires multiple tests and random chances. Through our study of the effects of spectral radius of the internal weight matrix on the training results, we propose to develop a method that can improve the echo-state network training by introducing a dynamic spectral radius. Our experiments verify that our new algorithm is significantly better than the original method for the training results and it is stable.