Table 2.
Deep generative data augmentation methods.
| Approach | References | Paradigm | Summary | |
|---|---|---|---|---|
| GAN | GAN (Goodfellow et al., 2014) |
Roy et al., 2020 | Motor imagery | Devised LSTM-based generator and discriminator; qualitatively analyzed generated signals |
| Krishna et al., 2020 | Speech | Devised GRU-based generator and discriminator | ||
| LSGAN (Mao et al., 2017) |
Pascual et al., 2019 | Seizure | Devised U-Net-based generator and discriminator; used conditional GAN concept | |
| DCGAN (Radford et al., 2015) |
Zhang et al., 2020b | Motor imagery | Generated STFT images estimated from raw EEGs; compared synthesizing quality to other DA methods | |
| Zhang and Liu, 2018 | Compared classification accuracy of testing dataset for different ratio of raw data and artificial data; used conditional GAN concept | |||
| Fahimi et al., 2020 | Motor | Used feature vector with the random noise for the generator input | ||
| Lee Y. E. et al., 2020 | ERP | Used features of EEG signals during walking as the generator input to reconstruct EEG signals similar to ones during standing | ||
| Truong et al., 2019a | Seizure | Generated STFT images estimated from raw EEGs | ||
| Truong et al., 2019b | Generated STFT images estimated from raw EEGs | |||
| Fan et al., 2020 | Sleep | Compared synthesizing quality to other DA methods | ||
| WGAN (Arjovsky et al., 2017) |
Ko et al., 2019 | Motor imagery | Conducted gradient penalty rather than weight clipping; used semi-supervised GAN concept | |
| Hartmann et al., 2018 | Motor | Conducted gradient penalty rather than weight clipping | ||
| Aznan et al., 2019 | SSVEP | Compared synthesizing quality to VAE-based DA methods; experimented TL setting | ||
| Panwar et al., 2019b | RSVP | Conducted gradient penalty rather than weight clipping; used conditional GAN concept | ||
| Luo et al., 2020 | Emotion | Conducted gradient penalty rather than weight clipping; used conditional GAN concept | ||
| Luo and Lu, 2018 | Conducted gradient penalty rather than weight clipping; used conditional GAN concept | |||
| Panwar et al., 2019a | Drowsy | Conducted gradient penalty rather than weight clipping | ||
| Hwang et al., 2019 | Cognition | Designed zero-calibration experiments | ||
| VAE | AE (Ballard, 1987) |
Zhang et al., 2020b | Motor imagery | Generated STFT images estimated from raw EEGs; compared synthesizing quality to other DA methods |
| VAE (Kingma and Welling, 2014) |
Zhang et al., 2020b | Motor imagery | Generated STFT images estimated from raw EEGs; compared synthesizing quality to other DA methods | |
| Fahimi et al., 2020 | Motor | Compared synthesizing quality to other DA methods | ||
| Aznan et al., 2019 | SSVEP | Compared synthesizing quality to VAE-based DA methods; experimented TL setting | ||
| Luo et al., 2020 | Emotion | Compared synthesizing quality to VAE-based DA methods |