A gradient-based automatic optimization CNN framework for EEG state recognition.

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Electroencephalogram (EEG) signals are crucial for assessing human psychophysiological states.
• Deep learning, particularly Convolutional Neural Networks (CNNs), excels at extracting features from EEG data.
• Designing effective CNNs for EEG analysis requires expertise and time, motivating automated approaches like Neural Architecture Search (NAS).

Purpose of the Study:

• To adapt and optimize an existing NAS algorithm (PC-DARTS) for the specific characteristics of EEG signals.
• To automate the design of CNN architectures for EEG-based classification tasks.
• To improve the efficiency and performance of EEG analysis using automated deep learning model design.

Main Methods:

• Modified the PC-DARTS algorithm to create a targeted search space for EEG signals.
• Incorporated frequency domain, time domain, and spatial electrode information into feature extraction.
• Applied the optimized NAS approach to EEG-based emotion recognition and driver drowsiness assessment.

Main Results:

• The novel NAS-derived CNN architectures achieved competitive accuracy on both emotion recognition and driver drowsiness tasks.
• The proposed method demonstrated a better standard deviation in performance across tasks compared to existing methods.
• The automated architecture design process proved effective for EEG signal analysis.

Conclusions:

• This work represents a successful application of NAS techniques to EEG data analysis.
• The optimized NAS approach offers a promising method for developing high-performance CNNs for various EEG classification and prediction tasks.
• This methodology can significantly reduce research time and broaden the application of CNNs in neuroscience and related fields.