Spiking Neural Network Decoder for Brain-Machine Interfaces.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Biomedical Engineering

Background:

• Neural prostheses decode brain activity to control external devices.
• Kalman filters are standard decoders for neural prosthetic applications, predicting movement trajectories.
• Spiking neural networks (SNNs) offer a biologically plausible alternative for neural data processing.

Purpose of the Study:

• To implement and evaluate a Kalman filter algorithm using a spiking neural network (SNN).
• To assess the performance of an SNN-based decoder for real-time neural prosthetic applications.
• To explore the potential of SNNs for efficient neural decoding in brain-computer interfaces.

Main Methods:

• Neural data from a rhesus monkey performing reaching movements were recorded using a 96-electrode array.
• A Kalman filter algorithm for predicting arm velocity was mapped onto an SNN using the Neural Engineering Framework.
• Simulations were performed using the Nengo software package.

Main Results:

• A 20,000-neuron SNN achieved decoding accuracy within 0.03% of the standard Kalman filter.
• A smaller 1,600-neuron SNN version operated in real-time with 0.27% error.
• The SNN implementation demonstrated comparable performance to the traditional Kalman filter.

Conclusions:

• Spiking neural networks can effectively implement complex signal processing algorithms like the Kalman filter for neural prostheses.
• SNNs can achieve high-performance neural decoding with significantly fewer neurons compared to traditional methods.
• Neuromorphic hardware implementations of SNNs hold promise for power-efficient, implantable neural prostheses.