[Paper Review] Robust and highly adaptable brain-computer interface with convolutional net architecture based on a generative model of neuromagnetic measurements.

Ivan Zubarev, Rasmus Zetter+2 more|arXiv (Cornell University)|May 28, 2018

EEG and Brain-Computer Interfaces3 citations

TL;DR

This paper proposes a convolutional neural network (CNN) architecture for non-invasive brain-computer interfaces (BCIs) using magnetoencephalographic (MEG) data, inspired by a generative model of neural signals for neurophysiological interpretability. The model achieves high accuracy in decoding event-related responses and oscillatory activity, generalizes across users without calibration, and enables incremental updates that boost real-time performance by 17.0%.

ABSTRACT

Deep Neural Networks have been applied very successfully in image recognition and natural language processing. Recently these powerful methods have received attention also in the brain-computer interface (BCI) community. Here, we introduce a convolutional neural network (CNN) architecture optimized for classification of brain states from non-invasive magnetoencephalographic (MEG) measurements. The model structure is motivated by a state-of-the-art generative model of the MEG signal and is thus readily interpretable in neurophysiological terms. We demonstrate that the proposed model is highly accurate in decoding event-related responses as well as modulations of oscillatory brain activity, and is robust with respect to inter-individual differences. Importantly, the model generalizes well across users: when trained on data pooled from previous users, it can successfully perform on new users. Thus, the time-consuming BCI calibration can be omitted. Moreover, the model can be incrementally updated, resulting in +8.9% average accuracy improvement in offline experiments and +17.0% in a real-time BCI. We argue that this model can be used in practical BCIs and basic neuroscience research.

Motivation & Objective

• Address the challenge of high inter-individual variability in non-invasive BCI systems by developing a model that generalizes across users.
• Reduce or eliminate the need for time-consuming user-specific calibration by leveraging transferable representations from pooled data.
• Design a deep learning architecture interpretable in neurophysiological terms by grounding it in a state-of-the-art generative model of MEG signals.
• Enable real-time BCI applications through incremental model updates that improve performance without retraining from scratch.

Proposed method

• Design a convolutional neural network (CNN) architecture that mirrors the structural and functional principles of a generative model of MEG signals, ensuring neurophysiological interpretability.
• Train the CNN on pooled MEG data from multiple subjects to enable cross-user generalization and eliminate individual calibration.
• Implement an incremental learning strategy that updates the model with new data from a new user without retraining on all prior data.
• Use spatial and temporal convolutional layers to extract features from MEG sensor time-series data, capturing both event-related responses and oscillatory modulations.
• Optimize the network using backpropagation with a cross-entropy loss function for binary or multi-class classification of brain states.
• Validate the model using both offline analysis and real-time BCI experiments to assess accuracy and adaptability.

Experimental results

Research questions

• RQ1 Can a deep learning model based on a generative MEG signal model achieve high decoding accuracy for event-related responses and oscillatory activity in non-invasive MEG data?
• RQ2 To what extent does the proposed CNN generalize across individuals without user-specific calibration?
• RQ3 How effective is incremental model updating in improving BCI performance over time with minimal data?
• RQ4 Does the neurophysiologically inspired architecture enhance interpretability while maintaining or improving classification accuracy?

Key findings

• The proposed CNN achieved high accuracy in decoding both event-related responses and modulations of oscillatory brain activity from MEG signals.
• The model generalized effectively across users: when trained on pooled data from previous subjects, it successfully decoded brain states in new users without individual calibration.
• Incremental model updates improved offline classification accuracy by 8.9% on average.
• In real-time BCI experiments, incremental updates led to a 17.0% average improvement in classification accuracy.
• The model’s architecture, grounded in a generative MEG model, provides neurophysiologically interpretable representations of brain activity.
• The system demonstrates strong robustness to inter-individual variability in MEG signal patterns.