Applying Neural Manifold Constraint on Point Process Model for Neural Spike Prediction

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computational Neuroscience

Background:

• Neural prostheses aim to restore function by modeling brain activity.
• Current models using point processes struggle with distorted single-neuron recordings.
• Neural population activity often lies on a lower-dimensional manifold.

Purpose of the Study:

• To develop a novel method for training neural prosthesis models that are robust to noisy recordings.
• To incorporate neural manifold properties into model training to improve prediction accuracy.
• To enhance the reliability of neural decoding for prosthetic applications.

Main Methods:

• Proposed a neural manifold constraint applied to the loss function during model training.
• The constraint minimizes the distance between model predictions and the empirical neural manifold.
• Utilized synthetic data with distorted spike trains for testing and validation.

Main Results:

• Models trained with the neural manifold constraint demonstrated a higher goodness-of-fit.
• The constraint effectively amended model predictions affected by distorted recordings.
• Kolmogorov-Smirnov test indicated improved similarity between predicted and original spike trains.

Conclusions:

• The neural manifold constraint offers a promising approach for improving neural prostheses.
• This method enhances model robustness in the presence of noisy neural data.
• The findings suggest potential for more reliable brain-computer interfaces in real-world applications.