FasterEEG: Adaptive Channel and Model Size Optimization for Efficient Brain Decoding.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Current electroencephalography (EEG) brain decoding methods prioritize accuracy over efficiency, limiting use in resource-limited settings.
• Fixed channel selection and model sizes in existing frameworks lead to computational redundancy and inflexibility.

Purpose of the Study:

• To develop a novel, differentiable framework (FasterEEG) for adaptive EEG brain decoding that optimizes both channel number and model size per sample.
• To enhance decoding efficiency and accuracy for practical brain-computer interface applications.

Main Methods:

• Proposed FasterEEG framework with an event-related desynchronization/synchronization (ERD/ERS)-based channel transformation for signal compression.
• Introduced a lightweight policy network for adaptive selection of input channels and model size.
• Employed a Gumbel-Estimator-based collaborative optimization for joint training of policy and classification networks.

Main Results:

• FasterEEG demonstrated consistent improvements in decoding efficiency across multiple datasets and baseline models.
• The framework maintained classification accuracy while significantly reducing computational cost.
• Achieved 1.89% higher accuracy with only 28% of the computational cost compared to ShallowConvNet.

Conclusions:

• FasterEEG offers a superior approach to EEG brain decoding by adaptively managing computational resources.
• The framework shows significant potential for accelerating brain-computer interfaces and enabling deployment in resource-constrained environments.
• Validated the feasibility and generalization ability of FasterEEG for efficient and accurate brain decoding.