Neural Network Based Comparison of Real and Synthetic Data Series in TeraHertz Domain

Abstract

Extension of real data by synthetic data becomes more important aspect of the virtualization technics today. In this paper we demonstrate how synthetic data generated from real data can be used in the supervised classification process of three different recurrent neural networks: Long-Short Term Memory (LSTM), Bidirectional LSTM (BiLSTM) and Gated Recurrent Unit (GRU). Other aspect is presented concerning the influence of the noise to the classification of real and synthetic data series. The paper demonstrates that LSTM network has better classification performance than GRU, even the last one has higher accuracy during the training. Synthetic data can eternalize just part of the features of the original real data and extraction efficiency of these characteristics depend on the applied neural network.