Low-power and lightweight spiking transformer for EEG-based auditory attention detection

Area of Science:

• Neuroscience
• Artificial Intelligence
• Signal Processing

Background:

• Electroencephalography (EEG) signal analysis aids in understanding brain activity, neural networks, and mechanisms of cognition, emotion, and behavior.
• EEG-based auditory attention detection infers attention to auditory stimuli by analyzing electrical brain activity.
• Current models face challenges in deployment on edge devices due to high computational demands.

Purpose of the Study:

• To develop a highly accurate, low-power, and lightweight model for EEG-based auditory attention detection suitable for edge devices.
• To overcome the computational limitations of traditional neural networks in EEG analysis.

Main Methods:

• A binarized spiking Transformer architecture was employed for EEG-based auditory attention detection.
• Spiking neurons were utilized to reduce power consumption through sparse, binary spike sequences.
• Post-training quantization converted full-precision weights to binary weights, minimizing model size.

Main Results:

• The binarized spiking Transformer demonstrated high accuracy in auditory attention detection.
• The model exhibited significantly reduced power consumption and a lightweight design.
• Experimental results showed performance exceeding state-of-the-art models with over a 21-fold reduction in model size.

Conclusions:

• The proposed binarized spiking Transformer is a pioneering solution for efficient EEG-based auditory attention detection on edge devices.
• The model offers a compelling balance of high accuracy, low power consumption, and reduced size.
• This approach facilitates more convenient and widespread application of EEG analysis in real-world scenarios.