Robustness of recurrent neural networks against deformation of external input patterns

This paper describes experimental investigations into the robustness of recurrent neural networks against deformation of external input patterns. Three types of deformation are prepared to demonstrate robustness of the networks: 1) superposition of Gaussian white noise; 2) nonlinear expansion and contraction along time axis; and 3) combination of the first with the second. The response of the network used shows that desired outputs are obtained from the deformed input patterns.<>