Personalized Continual EEG Decoding Framework for Knowledge Retention and Transfer

Dan Li,Hye-Bin Shin,Kang Yin

2024-11-04

Abstract:The significant inter-subject variability in electroencephalogram (EEG) signals often leads to knowledge being overwritten as new tasks are introduced in continual EEG decoding. While retraining on the entire dataset with each new input can prevent forgetting, this approach incurs high computational costs. An ideal brain-computer interface (BCI) model should continuously learn new information without retraining from scratch, thus reducing these costs. Most transfer learning models rely on large source-domain datasets for pre-training, yet data availability is frequently limited in real-world applications due to privacy concerns. Furthermore, such models are prone to catastrophic forgetting in continual EEG decoding tasks. To address these challenges, we propose a personalized subject-incremental learning (SIL) framework for continual EEG decoding that integrates Euclidean Alignment for fast domain adaptation, an exemplar replay mechanism to retain prior knowledge, and reservoir sampling-based memory management to handle memory constraints in long-term learning. Validated on the OpenBMI dataset with 54 subjects, our framework effectively balances knowledge retention with classification performance in continual MI-EEG tasks, offering a scalable solution for real-world BCI applications.

Signal Processing,Human-Computer Interaction

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is how to avoid knowledge from being overwritten or forgotten and reduce computational costs in continuous electroencephalogram (EEG) decoding tasks while continuously introducing new tasks. Specifically:

1. **Knowledge overwriting caused by cross - individual differences**: There are significant differences in EEG signals among different individuals, which makes the previously learned knowledge easily overwritten when new tasks are introduced.
2. **High computational cost of retraining**: In order to prevent knowledge forgetting, the method of retraining the entire dataset will lead to high computational costs.
3. **Data privacy and availability limitations**: Most transfer learning models rely on large - scale source - domain datasets for pre - training, but in practical applications, these data are often difficult to obtain due to privacy issues.
4. **Catastrophic Forgetting**: In the continuous learning process, the model is prone to forget the previously learned knowledge.

To solve these problems, the author proposes a personalized Subject - Incremental Learning (SIL) framework, which combines the following techniques:
- **Euclidean Alignment**: It is used for rapid domain adaptation and reducing domain differences among different individuals.
- **Exemplar Replay Mechanism**: By replaying previously learned samples, the existing knowledge is retained.
- **Reservoir Sampling - based Memory Management**: It deals with the memory limitation problem in long - term learning.

The experimental results show that this framework effectively balances knowledge retention and classification performance on the OpenBMI dataset and provides a scalable solution suitable for real - world brain - computer interface (BCI) applications.