Multi-class motor imagery EEG classification using collaborative representation-based semi-supervised extreme...

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalographic (EEG) data is crucial for brain-computer interfaces (BCI).
• Supervised learning methods struggle with limited labeled EEG data, while unlabeled data is abundant.
• Semi-supervised learning (SSL) leverages both data types but can be negatively impacted by unlabeled data.

Purpose of the Study:

• To develop a robust semi-supervised learning algorithm for EEG-based BCI that mitigates performance degradation from unlabeled data.
• To introduce a novel safety-control mechanism for evaluating the risk associated with unlabeled samples in SSL.

Main Methods:

• Proposed a collaborative representation-based semi-supervised extreme learning machine (CR-SSELM) algorithm.
• Implemented a safety-control mechanism to assess the risk of unlabeled samples using collaborative representation (CR).
• Integrated a risk-based regularization term into the semi-supervised extreme learning machine (SS-ELM) objective function.

Main Results:

• The CR-SSELM algorithm demonstrated superior performance compared to standard Extreme Learning Machine (ELM) and SS-ELM.
• In scenarios where SS-ELM underperformed ELM, CR-SSELM still achieved better results than ELM.
• Validated the effectiveness of the proposed safety mechanism in improving SSL performance on benchmark and EEG datasets.

Conclusions:

• The CR-SSELM algorithm effectively addresses the challenges of using unlabeled data in EEG-based BCI.
• The novel safety-control mechanism enhances the reliability and performance of semi-supervised learning.
• This approach offers a promising solution for improving BCI systems by maximizing the utility of available EEG data.