A Biologically Plausible Spiking Neural Network for Decoding Kinematics in the Hippocampus and Premotor Cortex

Abstract

This work presents a spiking neural network for predicting kinematics from neural data towards accurate and energy-efficient brain machine interface. A brain machine interface is a technological system that interprets neural signals to allow motor impaired patients to control prosthetic devices. Spiking neural networks have the potential to improve brain machine interface technology due to their low power cost and close similarity to biological neural structures. The SNN in this study uses the leaky integrate-and-fire model to simulate the behavior of neurons, and learns using a local learning method that uses surrogate gradient to learn the parameters of the network. The network implements a novel continuous time output encoding scheme that allows for regression-based learning. The SNN is trained and tested offline on neural and kinematic data recorded from the premotor cortex of a primate and the hippocampus of a rat. The model is evaluated by finding the correlation between the predicted kinematic data and true kinematic data, and achieves peak Pearson Correlation Coefficients of 0.77 for the premotor cortex recordings and 0.80 for the hippocampus recordings. The accuracy of the model is benchmarked against a Kalman filter decoder and a LSTM network, as well as a spiking neural network trained with backpropagation to compare the effects of local learning.

Benchmarking of hardware-efficient real-time neural decoding in brain–computer interfaces

“…Neural decoding tasks require metrics that incorporate the correction of the temporal regression to evaluate decoding accuracy. In such cases, the coefficient of determination (R 2 ) [3,11,20,30,[42][43][44] and the coefficient of correlation (Pearson's r) [42,[44][45][46][47] are widely used to assess the performance of neural decoding algorithms.…”

A Spiking Neural Network with Continuous Local Learning for Robust Online Brain Machine Interface

“…To enable 2D cursors control with a SNN, the SNN output must be modified to produce two continuous real-values functions corresponding to the x and y components of predicted arm velocity. SNN decoders can be adapted to produce these functions by adding a readout layer consisting of two leakyintegrate neurons that never emit spikes [20,23], where the output is given by the neuron's membrane potential. This method of computing the output has several desirable properties, such as the fact that the membrane potential is a continuous and differentiable function, so there will be no unnatural instantaneous changes in the predicted velocity.…”

“…The memory complexity of BPTT is O(N T ), where T is the input duration. In BMI applications, L is typically significantly less than T, since SNN performance don't increase significantly after the layer size increases beyond about 50-60 neurons [23], whereas the number of time steps could easily enter the hundreds of thousands for a continuous experiment that lasts longer than an hour, assuming a neural sampling rate of 100 Hz.…”

“…3.1.2. Batched Learning Comparison: SNN decoders have traditionally been trained using batched learning methods [20,23]. To ensure a valid comparison with existing decoders, the continuous learning model in this study was also trained using batched learning.…”

“…Previous work has explored the use of SNNs for neural decoding [20][21][22][23]. In [22], authors demonstrated an SNN that could replicate the behavior of a Kalman filter using the Neural Engineering Framework.…”