Improving The State Space Organization Of Untrained Recurrent Networks

Recurrent neural networks are frequently used in cognitive science community for modeling linguistic structures. More or less intensive training process is usually performed but several works showed that untrained recurrent networks initialized with small weights can be also successfully used for this type of tasks. In this work we demonstrate that the state space organization of untrained recurrent neural network can be significantly improved by choosing appropriate input representations. We experimentally support this notion on several linguistic time series.