Spiking Neural Network for Intra-cortical Brain Signal Decoding

TL;DR

This work tackles the need for accurate yet energy-efficient intra-cortical brain signal decoding for iBCIs by proposing a spiking neural network (SNN) architecture augmented with a feature fusion (FF) strategy that combines manually extracted NAV features with deep representations. The SNN employs a two-layer construct with temporal and spatial convolutions and PLIF neurons, trained with surrogate gradients to enable end-to-end learning. Experimental results on motor-related intra-cortical data from two rhesus macaques show that the SNN outperforms traditional artificial neural networks in decoding accuracy while delivering orders of magnitude improvements in energy efficiency, with FF providing additional gains. The proposed approach thus offers a practical path toward high-precision, low-power iBCIs suitable for real-time neural decoding on neuromorphic hardware.

Abstract

Decoding brain signals accurately and efficiently is crucial for intra-cortical brain-computer interfaces. Traditional decoding approaches based on neural activity vector features suffer from low accuracy, whereas deep learning based approaches have high computational cost. To improve both the decoding accuracy and efficiency, this paper proposes a spiking neural network (SNN) for effective and energy-efficient intra-cortical brain signal decoding. We also propose a feature fusion approach, which integrates the manually extracted neural activity vector features with those extracted by a deep neural network, to further improve the decoding accuracy. Experiments in decoding motor-related intra-cortical brain signals of two rhesus macaques demonstrated that our SNN model achieved higher accuracy than traditional artificial neural networks; more importantly, it was tens or hundreds of times more efficient. The SNN model is very suitable for high precision and low power applications like intra-cortical brain-computer interfaces.

Figures (9)

• Figure 1: A closed-loop BCI system.
• Figure 2: (a) The computational process of a PLIF neuron. (b) The Heaviside function and two surrogate functions, and their derivatives.
• Figure 3: Architecture of our proposed SNN. The first convolution layer performs channel-wise temporal convolution, with one convolution kernel for each channel. The second layer is spatial convolution to extract the spatial information across different channels. The third layer is a fully connected layer. In each convolution layer, a batch normalization (BN) layer and a PLIF neuron is employed after the kernel.
• Figure 4: Feature fusion. We construct a deep neural network (ANN or SNN) and remove its classifier to extract the deep representations. They and the hand-crafted NAV features are then separately and linearly projected into a hidden space. The two projections are concatenated as input to the linear classifier.
• Figure 5: The neural signal recording settings. (a) The reaching task: Three cylindrical objects were placed on a panel, with a target light positioned above each object. In each trial, the monkey performed a reaching movement guided by the illumination of a specific target light. (b) The grasping task: One of three different shapes (cube, triangle, sphere) was presented at the panel's center via a motor, for the monkey to grasp.