Decoding Speech from Single Trial MEG Signals Using Convolutional Neural Networks and Transfer Learning.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) offer communication solutions for patients with severe motor impairments, such as locked-in syndrome.
• Decoding speech directly from neural signals is a key goal for advancing BCI technology.
• Magnetoencephalography (MEG) provides rich temporal and spectral information about brain activity.

Purpose of the Study:

• To investigate the use of spectral and temporal features from MEG signals for speech decoding.
• To train convolutional neural networks (CNNs) for classifying neural signals related to spoken phrases.
• To enhance the efficiency of CNN training for BCIs using dimensionality reduction and transfer learning.

Main Methods:

• Extracted spectral and temporal features from magnetoencephalography (MEG) data.
• Trained convolutional neural networks (CNNs) to classify neural signals corresponding to perceived, imagined, and produced speech.
• Applied principal component analysis (PCA) for spatial dimension reduction and transfer learning for model initialization to accelerate training.

Main Results:

• CNNs effectively decoded speech signals across perception, imagination, and production tasks.
• Principal component analysis (PCA) and transfer learning significantly reduced CNN training time (over 10x faster).
• The optimized approach (50 principal coefficients + transfer learning) maintained high speech decoding accuracy.

Conclusions:

• CNNs are effective for decoding speech from MEG signals, enabling advanced BCIs.
• PCA and transfer learning are crucial for improving the practical usability of CNN-based BCI systems by reducing training time.
• This research paves the way for more efficient communication tools for individuals with communication disabilities.