The researchers Zhenis Otarbay and Abzal Kyzyrkanov developed ConvoReleNet, a neural network combining convolutional layers, Transformer-based attention and bidirectional long short-term memory layers. The model learns spatial and temporal EEG patterns and transfers knowledge acquired from multiple participants to previously unseen users.

The framework was tested on two public BCI Competition IV datasets. BNCI IV-2a includes four motor-imagery classes recorded from 22 EEG channels, while BNCI IV-2b includes left- and right-hand imagery recorded from three channels.

Without transfer learning, the model achieved an average accuracy of 72.22% on IV-2a and 75.10% on IV-2b. Transfer learning increased accuracy to 79.44% and 83.85%, respectively.

It also reduced differences in performance between participants. Across-subject variation decreased by 45.9% on IV-2a and 40% on IV-2b.

The highest accuracy on IV-2a was 87.55%, achieved using the Tanh activation function. On IV-2b, the best result was 83.85% with ELU activation.

The study also showed that conservative fine-tuning was more effective than aggressive retraining. Reducing the learning rate and number of training epochs increased accuracy on IV-2b from 74.36% to 77.64%. This approach helped the model adapt to new data without losing useful knowledge acquired during pretraining.

The compact transfer-learning configuration contained approximately 1.38 million trainable parameters, providing a balance between accuracy and computational efficiency.

The results indicate that transfer learning can reduce the need to train a separate BCI model from scratch for every user. Further validation is required using larger datasets, clinical populations and recordings collected in real-world conditions.