Brain-inspired modular echo state network for EEG-based emotion recognition.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Affective Computing

Background:

• Recurrent Neural Networks (RNNs), specifically Echo State Networks (ESNs), are used for EEG-based emotion recognition.
• Existing methods using intrinsic plasticity (IP) and synaptic plasticity (SP) for ESN tuning are computationally complex and require extra training.
• Neuroscience shows that modular brain network topology enhances information processing efficiency.

Purpose of the Study:

• To propose a novel Modular Echo State Network (M-ESN) by initializing the ESN hidden layer with a modular structure.
• To develop an implementation method for optimizing module numbers and connectivity in M-ESN.
• To evaluate the performance of M-ESN on EEG emotion recognition tasks and understand the benefits of modularity.

Main Methods:

• Developed a novel implementation for initializing ESNs with a modular structure (M-ESN).
• Determined optimal module numbers and local/global connectivity parameters.
• Benchmarked M-ESN performance on the DEAP dataset for emotion recognition tasks.

Main Results:

• M-ESN significantly outperforms regular ESN in classifying emotion arousal (5.44%), valence (5.90%), and stress/calm (5.42%).
• M-ESN shows competitive or superior performance compared to adaptation rule-tuned ESNs (0.77-5.49% difference).
• Achieved these improvements with a smaller reservoir size and a simpler training process.

Conclusions:

• Network modularity in M-ESN leads to a more diverse distribution of node degrees, enhancing network heterogeneity.
• This heterogeneity improves classification accuracy in EEG-based emotion recognition.
• M-ESN offers a more efficient and computationally simpler alternative for affective computing applications.