Proposed architecture (still in testing):
- Hybrid Network (CNN + LSTM-RNN)
- Extract EEG signals based on time frame and map them into images
- LSTM is used to preserve the time dimension
- Transfer Learning
- Target experiment data (from children with speech deficiencies) is limited, and come from different domains as the training dataset
- Assuming emotions are influenced by individual experiences, by TL we could minimise individual parity