Abstract
During the last decade, deep learning (DL) techniques have demonstrated the capabilities in various applications with a large number of labeled samples. Unfortunately, it is normally difficult to obtain such large amounts of samples in practice. As one of the most promising research directions of data generation in DL, generative adversarial network (GAN) can process not only images but also time-series signals. Unfortunately, it is easy to lose the time-dependence information from the latter due to characteristics of GAN, which increases the challenge of signal generation. Besides, the existing evaluation methods cannot evaluate the performance of GAN comprehensively. Therefore, this paper summarizes the current work of time-series signals generation based on GAN and the existing evaluation methods of GAN. As compared to existing GAN-related review work, this paper claims four unique points: (1) we specify the difficulties of GAN for time-series generation, particularly for the biological signal generation with potential solutions; (2) we analyze drawbacks of existing evaluation methods, and propose feasible solutions; (3) some suggestions are provided for the further research of robust time-series signal generation, especially for biological signal generation; (4) we provide a preliminary experiment to demonstrate the effectiveness of GANs for time-series signals generation, particularly, electroencephalogram (EEG).