Motor decoding from the posterior parietal cortex using deep neural networks

Area of Science:

• Neuroscience
• Machine Learning
• Brain-Computer Interfaces (BCIs)

Background:

• Motor decoding translates neural activity for BCIs, offering insights into brain encoding of motor states.
• Deep Neural Networks (DNNs) are emerging as powerful tools for neural decoding.
• The optimal DNN for various motor decoding challenges and invasive BCIs remains unclear.

Purpose of the Study:

• To compare the performance of different DNNs (FCNNs, CNNs, RNNs) for motor decoding.
• To evaluate DNNs across diverse motor tasks, including reaching and reach-to-grasping.
• To assess DNN robustness under reduced neuron/trial counts and explore transfer learning potential.

Main Methods:

• Designed and applied Fully-Connected Neural Networks (FCNNs), Convolutional Neural Networks (CNNs), and Recurrent Neural Networks (RNNs).
• Decoded motor states from V6A posterior parietal cortex neuronal recordings in macaques during three motor tasks.
• Analyzed performance with reduced data, employed transfer learning, and examined accuracy time courses for motor encoding insights.

Main Results:

• DNNs surpassed traditional Naïve Bayes classifiers; CNNs outperformed XGBoost and Support Vector Machines.
• CNNs demonstrated superior performance with fewer neurons and trials, and transfer learning enhanced decoding in low-data scenarios.
• V6A neurons encoded reaching and grasp properties, with grip encoding occurring later and being weaker in darkness.

Conclusions:

• CNNs are effective candidates for human invasive BCIs using posterior parietal cortex recordings.
• CNNs can reduce BCI calibration times through transfer learning.
• CNN-based analysis offers insights into neural encoding properties and brain region functions.