Table 4.
Implicit transfer learning methods.
| Approach | References | Paradigm | Summary | |
|---|---|---|---|---|
| Fine-tuning | Whole | Shovon et al., 2019 | Motor imagery | Pre-trained with natural images |
| Aznan et al., 2019 | SSVEP | Pre-trained with synthetic SSVEP samples | ||
| Andreotti et al., 2018 | Sleep | Trained their network with source subjects and fine-tuned it with target subject (LOO) | ||
| Phan et al., 2020 | Pre-trained network with different dataset | |||
| Vilamala et al., 2017 | Pre-trained network with natural images | |||
| Fahimi et al., 2019 | Cognition | Trained their network with source subjects and fine-tuned it with target subject (LOO) | ||
| Partial | Zhang et al., 2021 | Motor imagery | Fine-tuned only fully-connected layers | |
| Zhao et al., 2019 | Conducted ablation studies to identify which layer should be transferred target | |||
| Raghu et al., 2020 | Seizure | Fine-tuned the last some layers of pre-trained network | ||
| Olesen et al., 2020 | Sleep | Fine-tuned parts of parameters | ||
| Enhancing representational power |
Attention | Zhang et al., 2018a | Motor imagery | Designed a self-attention module to find more class-discriminative segments |
| Zhang et al., 2019a | Designed a recurrent self-attention module | |||
| Zhang et al., 2020a | Presented raw EEG to a spatial graph and designed a recurrent self-attention module | |||
| Zhang et al., 2019b | Presented raw EEG to a spatial graph and designed two attention modules; one for attentive temporal dynamics and the other for attentive channels | |||
| Multi-scale features | Kwon et al., 2019 | Extracted spatio-spectral features in multi-frequency bands using CSP and selected top bands to use them as inputs | ||
| Ko et al., 2020a | Multi | Extracted multi-scale features including spatio-temporal-spectral patterns | ||
| Maximize mutual information | Jeon et al., 2020 | Motor imagery | Decomposed an intermediate feature into a class-relevant and class-irrelevant feature and maximized mutual information between low-level and high-level representations | |
| Meta-learning | MAML (Finn et al., 2017) | Duan et al., 2020 | Multi | Trained optimal parameters through gradient-based optimization and conducted fine-tuning with a small amount of target data |
| Relation (Sung et al., 2018) | An et al., 2020 | Motor imagery | Estimated relation scores between support and query sets among source subjects in few-shot scenarios |