Time Warping Invariant Echo State Networks

Echo State Networks (ESNs) is a recent simple and powerful approach to training recurrent neural networks (RNNs). In this report we present a modification of ESNs - time warping invariant echo state networks (TWIESNs) that can effectively deal with time warping in dynamic pattern recognition. The standard approach to classify time warped input signals is to align them to candidate pro- totype patterns by a dynamic programming method and use the alignment cost as a classification criterion. In contrast, we feed the original input signal into specifically designed ESNs which intrinsically are invariant to time warping in the input. For this purpose, ESNs with leaky integrator neurons are required, which are here presented for the first time, too. We then explain the TWIESN architecture and demonstrate their functioning on very strongly warped, synthetic data sets.