Machine to brain: facial expression recognition using brain machine generative adversarial networks.

Area of Science:

• Artificial Intelligence
• Neuroscience
• Computer Vision

Background:

• Deep neural networks for facial expression recognition (FER) are data-dependent and lack human-like cognitive abilities.
• The human brain performs FER effectively with few samples by leveraging cognitive processing.
• Bridging the gap between AI and human cognitive capabilities in FER is a significant challenge.

Purpose of the Study:

• To propose a novel framework, Brain Machine Generative Adversarial Networks (BM-GAN), to improve FER performance.
• To enable deep neural networks to generate LIKE-electroencephalograph (EEG) features, mimicking human cognitive processes.
• To achieve human-like performance in FER by integrating visual and cognitive feature generation.

Main Methods:

• Obtaining EEG signals triggered from facial emotion images.
• Utilizing BM-GAN for the mutual generation of image visual features and EEG cognitive features.
• Developing VisualNet for image feature extraction and EEGNet for cognitive feature extraction.

Main Results:

• Achieved an average classification accuracy of 96.6% on the Chinese Facial Affective Picture System dataset.
• Demonstrated successful FER using LIKE-EEG features without the need for EEG signals during the testing phase.
• Validated the effectiveness of BM-GAN in generating cognitive features that enhance FER.

Conclusions:

• The proposed BM-GAN framework significantly enhances FER performance by incorporating cognitive principles.
• LIKE-EEG features generated by BM-GAN enable accurate FER, comparable to human capabilities.
• The method offers a promising direction for developing more cognitively adept AI systems for complex tasks like FER.