Simplified long short-term memory model for robust and fast prediction

摘要

Long short-term memory(LSTM) is an effective solution to time sequence prediction. Considering the data perturbations, in this letter, a variant model of LSTM is proposed to achieve robustness of prediction. Specifically, data processing procedure in the recurrent unit of proposed model is reformulated, the gates are controlled by only one variable, and the variable is the sum of long-term memory and the current input. Due to the simplified two-gate structure of proposed model, the speed of prediction is improved as well. The experiments on three datasets verify that the proposed model with simplified structure has higher robustness and shorter running time than the traditional LSTM model.