Application of the accelerated PSO method to training LSTM recurrent neural networks

Source abstract

The method of training based on accelerated particle swarm optimization (APSO) for recurrent neural networks (LSTM - long short-term memory) is considered. The paper examines the algorithm for weight correction and the efficiency of LSTM network training. In recurrent neural networks (RNN), there is the problem of vanishing gradient between early and later neural network layers during long-term forecasting. The proposed method of applying APSO to train LSTM-RNN addresses this problem and leads to the creation of a more efficient recurrent neural network for long-term forecasting. An example of training LSTM for ECG time series using the APSO method is provided.