A subspace approach to learning recurrent features from brain activity.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain signal recordings exhibit significant instability and time variability across sessions.
• This variability stems from physiological changes, learning effects, and environmental factors.
• Such signal degradation poses a major challenge for reliable brain-computer interfaces.

Purpose of the Study:

• To develop a novel technique for addressing signal instability in neural recordings.
• To improve classification accuracy in brain-machine interfaces (BMIs) by leveraging time-recurrent neural activity.
• To enhance movement direction decoding using local field potential (LFP) data.

Main Methods:

• Proposed a method to learn subspaces that are recurrent over time from neural data.
• Utilized projections onto these learned subspaces for signal classification.
• Applied the technique to movement direction decoding using local field potential (LFP) signals.

Main Results:

• Demonstrated significant improvement in decoding power, increasing accuracy from 76% to 88% for an unstable subject.
• Showcased consistent decoding performance across different subjects.
• Validated the effectiveness of recurrent subspace learning for stabilizing neural signals.

Conclusions:

• The proposed technique effectively mitigates time variability in neural recordings.
• Learning recurrent subspaces offers a robust approach for enhancing BMI performance.
• This method holds promise for more reliable and accurate neural decoding applications.