LSTM-Based Online Learning with Extended Kalman Filter Based Training Algorithm

In this study, we work on online learning with long short term memory (LSTM) based networks is investigated. For LSTM-based online learning, a computationally efficient training algorithm based on extended Kalman filter (EKF) is proposed. We show the considerable performance improvements achieved by the proposed algorithm through comparing with conventional LSTM methods in simulations. Results particularly show that the proposed algorithm achieves very similar performance with conventional methods in 25-40 times shorter training time depending on the network complexity.