M1M2: Deep-Learning-Based Real-Time Emotion Recognition from Neural Activity.

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Emotion recognition is crucial for human-computer interaction, but image-based methods are unreliable due to intentional expression masking.
• Electroencephalography (EEG) signals offer a more objective measure of emotions, yet classifying them remains challenging for current machine learning and deep learning techniques.

Purpose of the Study:

• To develop highly accurate emotion recognition models using EEG signals.
• To address the limitations of existing machine learning and deep learning approaches for EEG signal classification.
• To propose novel convolutional neural network (CNN) architectures for effective emotion detection.

Main Methods:

• Two CNN models (M1: heavily parameterized, M2: lightly parameterized) were developed and combined with advanced feature extraction techniques.
• Fast Fourier Transformation was employed for frequency domain feature extraction, complemented by deep features from convolutional layers.
• The DEAP dataset, a popular benchmark for EEG analysis, was used for binary classification of valence and arousal.

Main Results:

• The M1 and M2 CNN models achieved exceptional accuracies of 99.89% and 99.22%, respectively, surpassing all prior state-of-the-art models.
• The M2 model demonstrated remarkable efficiency, achieving 99.22% accuracy with just 2 seconds of EEG data and over 96% accuracy with only 125 milliseconds for valence classification.
• The M2 model also showed high performance with limited data, reaching 96.8% accuracy on valence using only 10% of the training dataset.

Conclusions:

• The proposed CNN models, particularly M2, offer a highly effective and efficient solution for emotion recognition from EEG signals.
• The study highlights the potential of deep learning, specifically CNNs, to overcome the challenges of EEG signal classification for affective computing.
• Reproducibility is ensured through the public release of documented implementation codes for all experiments.