Analysis of Gradient Vanishing of RNNs and Performance Comparison

Abstract

A recurrent neural network (RNN) combines variable-length input data with a hidden state that depends on previous time steps to generate output data. RNNs have been widely used in time-series data analysis, and various RNN algorithms have been proposed, such as the standard RNN, long short-term memory (LSTM), and gated recurrent units (GRUs). In particular, it has been experimentally proven that LSTM and GRU have higher validation accuracy and prediction accuracy than the standard RNN. The learning ability is a measure of the effectiveness of gradient of error information that would be backpropagated. This study provided a theoretical and experimental basis for the result that LSTM and GRU have more efficient gradient descent than the standard RNN by analyzing and experimenting the gradient vanishing of the standard RNN, LSTM, and GRU. As a result, LSTM and GRU are robust to the degradation of gradient descent even when LSTM and GRU learn long-range input data, which means that the learning ability of LSTM and GRU is greater than standard RNN when learning long-range input data. Therefore, LSTM and GRU have higher validation accuracy and prediction accuracy than the standard RNN. In addition, it was verified whether the experimental results of river-level prediction models, solar power generation prediction models, and speech signal models using the standard RNN, LSTM, and GRUs are consistent with the analysis results of gradient vanishing.