Unsupervised representation learning of spontaneous MEG data with nonlinear ICA

Area of Science:

• Neuroscience
• Computational Neuroscience
• Signal Processing

Background:

• Resting-state magnetoencephalography (MEG) reveals complex spatiotemporal brain activity patterns.
• The neurophysiological underpinnings and source separation of these signals remain challenging.
• Existing methods often struggle with the inherent complexity of neural data.

Purpose of the Study:

• To develop an unsupervised learning method for representing spontaneous cortical activity from MEG data.
• To investigate the utility of nonlinear independent component analysis (ICA) for analyzing complex neural signals.
• To assess the performance of nonlinear ICA in downstream tasks with limited labeled data.

Main Methods:

• Implemented a nonlinear independent component analysis (ICA) model, a generative approach trained via unsupervised learning.
• Trained the model on a large resting-state MEG dataset from the Cam-CAN repository.
• Applied the learned representations to audio-visual classification and attentional state decoding tasks.

Main Results:

• The nonlinear ICA model learned to represent and generate spontaneous cortical activity patterns using latent nonlinear components.
• Achieved competitive performance in audio-visual MEG classification compared to deep neural networks with limited labels.
• Demonstrated generalizability by decoding attentional states from a neurofeedback dataset with ~70% individual accuracy, surpassing linear ICA and baseline methods.

Conclusions:

• Nonlinear ICA is a powerful tool for unsupervised representation learning of spontaneous MEG activity.
• This method offers a valuable approach for analyzing complex neural data, especially when labeled datasets are scarce.
• The findings highlight the potential of nonlinear ICA for real-time feature extraction and decoding in various neuroscience applications.