Comparison of LSTM and GRU Recurrent Neural Network Architectures

This paper describes the comparison results of two types of recurrent neural network: LSTM and GRU. In the article the two types of RNN architecture are compared with the criterion of time consumed for test problems solving and training. Information about network training is provided in order to explain the differences in the training of LSTM and GRU RNN’s types and the final difference in time. Mathematic models of this neural network types are provided. The article includes description of software implementation of recurrent neural networks. As a result of research the numerical comparison of training and solving time is provided, and practical hints and conclusions are derived.