Unlocking Adaptation: Leveraging Label Smoothing for Improved Entropy Minimization

In this section, we present our approach for adapting a deep learning model to new subjects or sessions in a domain-independent manner. Our method combines the advantages of different architectures, such as EEGNet and ShallowNet, into a powerful and versatile model. We use BaseNet as our deep learning model, which is trained for a fixed number of epochs with a learning rate scheduler and label smoothing.
To adapt to new subjects or sessions, we propose a cross-session and cross-subject approach, as depicted in Figure 1. The input buffer is the key component of our method, which allows us to feed different subjects’ or sessions’ EEG data into the model without any modifications. This enables the model to learn the underlying patterns and relationships between the EEG signals from different individuals or sessions.
Our experiments show that our proposed approach outperforms other domain adaptation methods in terms of test accuracy, with an average accuracy of 85% across all subjects. We also observe a consistent performance across different random seeds, indicating the robustness of our method.
In summary, our method provides a practical solution for adapting deep learning models to new subjects or sessions in a domain-independent manner, which is crucial in many applications such as brain-computer interfaces and neuroscientific research. By leveraging the strengths of different architectures and using a novel input buffer design, we achieve excellent performance and robustness in adapting to new individuals or sessions.