Spatio-temporal transformers for decoding neural movement control.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Neuroscience

Background:

• Deep learning advances neurophysiological data analysis, but designing interpretable artificial neural networks for in vivo neural activity remains challenging.
• Balancing efficiency in low-data conditions with result interpretability is crucial for practical applications in neural decoding.

Purpose of the Study:

• To develop and evaluate a specialized transformer architecture for analyzing single-neuron spiking activity.
• To assess the model's predictive capabilities in motor control tasks and its interpretability.

Main Methods:

• A novel specialized transformer architecture was designed to analyze single-neuron spiking activity.
• The model was tested on multi-electrode recordings from the dorsal premotor cortex of non-human primates during a motor inhibition task.

Main Results:

• The architecture achieved early prediction of movement direction (within 230 ms of the Go signal).
• The model successfully forecasted movement generation or withholding before a stop signal presentation.
• Analysis of internal model dynamics, including predicted correlations, mirrored previous theoretical findings.

Conclusions:

• The proposed framework demonstrates a practical application of deep learning in motor control research.
• The architecture offers both predictive power and interpretability for analyzing neural activity.
• This work advances the use of artificial intelligence in understanding neural processes underlying motor behavior.