Decoding and interpreting cortical signals with a compact convolutional neural network.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) translate neural activity into commands for external devices.
• Deep learning models automate feature extraction in BCIs, but interpreting their parameters for physiological insights remains challenging.

Purpose of the Study:

• To introduce a compact convolutional neural network (CNN) architecture for adaptive decoding of electrocorticographic (ECoG) data into finger kinematics.
• To propose a theoretically justified method for interpreting spatial and temporal weights in CNNs for BCI applications.
• To enable automatic knowledge discovery from neural data by interpreting network parameters.

Main Methods:

• Developed a compact CNN architecture for adaptive decoding of ECoG data.
• Proposed a novel approach for interpreting spatial and temporal weights in the CNN architecture.
• Validated the approach using Monte Carlo simulations and real ECoG and EEG datasets.

Main Results:

• The CNN architecture achieved performance comparable to competition winners on ECoG data without manual feature engineering.
• The interpretation method revealed spatial and spectral patterns of neural processes underlying finger kinematics decoding.
• Physiologically plausible patterns were observed when applying the pipeline to EEG motor-imagery data.

Conclusions:

• The proposed CNN architecture is compact, interpretable, and effective for BCI decoding.
• The theoretically justified weight interpretation rules offer a novel tool for investigating neural mechanisms of motor control.
• This work advances BCI technology by combining robust decoding with physiological insight.