AutoEER: automatic EEG-based emotion recognition with neural architecture search.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Emotion recognition using electroencephalography (EEG) is increasingly important for applications with portable devices.
• Deep learning models excel at EEG emotion recognition but require time-consuming manual design and customization.
• Neural architecture search (NAS) offers automated solutions for optimizing deep networks.

Purpose of the Study:

• To introduce AutoEER, a framework using tailored NAS for automatic optimal network structure discovery in EEG-based emotion recognition.
• To design a specialized search space capturing temporal and spatial EEG properties.
• To develop a novel parameterization strategy for deriving optimal network structures.

Main Methods:

• Proposed AutoEER framework leveraging NAS for EEG emotion recognition.
• Developed a customized search space incorporating operators for temporal and spatial EEG features.
• Implemented a novel parameterization strategy for network structure optimization.

Main Results:

• AutoEER outperformed state-of-the-art manual and NAS models on DEAP and SEED datasets.
• Achieved a 0.93% average accuracy improvement over WangNAS.
• Achieved a 4.51% average F1 score improvement over LiNAS.
• Generated architectures demonstrated superior transferability.

Conclusions:

• AutoEER offers a novel, automated approach to EEG-based emotion recognition model design.
• The specialized search space and parameterization strategy yield high-performing, transferable models.
• AutoEER significantly reduces manual labor and time costs in EEG research, promising to advance the field.