Grammatical Structure Detection with Intrinsic Plasticity based Echo State Networks for Cognitive Robot

摘要

A novel model called Intrinsic plasticity Echo State Network with new weight initialization (IP-ESN) is proposed in this study. The new input weight, which is selected by proposed method called new weight initialization, replaces the input weight of Echo State Network (ESN). Intrinsic plasticity rules is applied to find the gain and bias before training model. These two elements can be used to extend the neurons connection of ESN, which directly impacts on the model performance. According to the results of experiments, comparing with ESN, new weight initialization and intrinsic plasticity rules play a vital roles in decreasing the error rate. Moreover, IP-ESN has the super understanding ability of meaning words and sentence in the different construction numbers of corpus than other compared models. This algorithm also can be applied in human robot applications, because the proposed algorithm can understand the meaning of sentence and predict the grammatical structure in the real time.