MOCNN: A Multiscale Deep Convolutional Neural Network for ERP-Based Brain-Computer Interfaces.

Area of Science:

• Neuroscience and Artificial Intelligence
• Brain-Computer Interfaces (BCIs)
• Signal Processing and Machine Learning

Background:

• Event-related potentials (ERPs) are crucial neurophysiological signals reflecting brain responses to external stimuli.
• Extracting complex spatiotemporal features from ERPs is essential for understanding brain activity.
• Deep learning models offer advanced capabilities for analyzing oscillatory activity in ERPs for BCIs.

Purpose of the Study:

• To develop a novel deep learning architecture for improved ERP classification in BCIs.
• To effectively mine discriminative spatiotemporal features from ERP signals across multiple frequencies.
• To introduce and apply the concept of octave convolution for multiscale feature extraction in ERP-BCI research.

Main Methods:

• Proposed a multiscale feature fusion octave convolution neural network (MOCNN) model.
• MOCNN processes ERP signals by dividing them into high-, medium-, and low-frequency components across different network branches.
• Employed temporal and spatial convolutions for feature mapping and utilized information exchange between branches for interactive learning.

Main Results:

• The MOCNN model demonstrated state-of-the-art performance in ERP classification.
• Achieved superior classification accuracy on two public datasets and one self-collected ERP dataset.
• The multiscale approach enriched feature information and optimized calculations by incorporating lower-frequency components.

Conclusions:

• MOCNN effectively extracts discriminative spatiotemporal features from multiscale ERP data.
• The introduction of octave convolution into ERP-BCI research enables advanced feature extraction.
• The proposed model significantly advances the performance of ERP-based brain-computer interfaces.