Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Effective and Efficient Intracortical Brain Signal Decoding with Spiking Neural Networks

Haotian Fu, Peng Zhang, Song Yang, Herui Zhang, Ziwei Wang, Dongrui Wu

TL;DR

The paper tackles the challenge of achieving high decoding accuracy while minimizing energy consumption in invasive brain–computer interfaces by introducing LSS-CA-SNN, an SNN architecture that combines local synaptic stabilization and channel-wise attention. It is augmented with SpikeDrop, a spiking data augmentation strategy, to improve generalization on intracortical spike data. Across two rhesus macaque datasets and two transfer scenarios, the approach outperforms state-of-the-art ANN baselines in accuracy and achieves substantial energy savings, demonstrating the practicality of energy-efficient, spike-based decoding for invasive BCIs. Together, these contributions advance the feasibility of real-time, low-power invasive BCIs for applications requiring high precision and longevity on neuromorphic hardware.

Abstract

A brain-computer interface (BCI) facilitates direct interaction between the brain and external devices. To concurrently achieve high decoding accuracy and low energy consumption in invasive BCIs, we propose a novel spiking neural network (SNN) framework incorporating local synaptic stabilization (LSS) and channel-wise attention (CA), termed LSS-CA-SNN. LSS optimizes neuronal membrane potential dynamics, boosting classification performance, while CA refines neuronal activation, effectively reducing energy consumption. Furthermore, we introduce SpikeDrop, a data augmentation strategy designed to expand the training dataset thus enhancing model generalizability. Experiments on invasive spiking datasets recorded from two rhesus macaques demonstrated that LSS-CA-SNN surpassed state-of-the-art artificial neural networks (ANNs) in both decoding accuracy and energy efficiency, achieving 0.80-3.87% performance gains and 14.78-43.86 times energy saving. This study highlights the potential of LSS-CA-SNN and SpikeDrop in advancing invasive BCI applications.

Effective and Efficient Intracortical Brain Signal Decoding with Spiking Neural Networks

TL;DR

The paper tackles the challenge of achieving high decoding accuracy while minimizing energy consumption in invasive brain–computer interfaces by introducing LSS-CA-SNN, an SNN architecture that combines local synaptic stabilization and channel-wise attention. It is augmented with SpikeDrop, a spiking data augmentation strategy, to improve generalization on intracortical spike data. Across two rhesus macaque datasets and two transfer scenarios, the approach outperforms state-of-the-art ANN baselines in accuracy and achieves substantial energy savings, demonstrating the practicality of energy-efficient, spike-based decoding for invasive BCIs. Together, these contributions advance the feasibility of real-time, low-power invasive BCIs for applications requiring high precision and longevity on neuromorphic hardware.

Abstract

A brain-computer interface (BCI) facilitates direct interaction between the brain and external devices. To concurrently achieve high decoding accuracy and low energy consumption in invasive BCIs, we propose a novel spiking neural network (SNN) framework incorporating local synaptic stabilization (LSS) and channel-wise attention (CA), termed LSS-CA-SNN. LSS optimizes neuronal membrane potential dynamics, boosting classification performance, while CA refines neuronal activation, effectively reducing energy consumption. Furthermore, we introduce SpikeDrop, a data augmentation strategy designed to expand the training dataset thus enhancing model generalizability. Experiments on invasive spiking datasets recorded from two rhesus macaques demonstrated that LSS-CA-SNN surpassed state-of-the-art artificial neural networks (ANNs) in both decoding accuracy and energy efficiency, achieving 0.80-3.87% performance gains and 14.78-43.86 times energy saving. This study highlights the potential of LSS-CA-SNN and SpikeDrop in advancing invasive BCI applications.
Paper Structure (20 sections, 10 equations, 7 figures, 7 tables)

This paper contains 20 sections, 10 equations, 7 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. ANNs for Non-Invasive BCI Decoding
  4. SNNs for Invasive BCI Decoding
  5. Data Augmentation
  6. Methods
  7. Preliminaries of SNN
  8. SNN with Synaptic Stabilization and Attention
  9. Spatiotemporal feature extraction
  10. Channel fusion with multilayer spatiotemporal convolutions
  11. Classifier
  12. SpikeDrop Data Augmentation
  13. Experiments and Results
  14. Dataset Description
  15. Experiment Settings
  16. ...and 5 more sections

Figures (7)

  • Figure 1: A closed-loop BCI system.
  • Figure 2: Our proposed LSS-CA-SNN, which includes four different convolutional layers to extract both spatial and temporal features from spiking signals.
  • Figure 3: A convolution-based PLIF-SNN layer, including convolution, membrane potential integration, spiking activity, and synaptic leak.
  • Figure 4: SpikeDrop augmentation process, showcasing the application of masks from various perspectives to manipulate the spiking data.
  • Figure 5: The data collection device in (a) motor paradigm and (b) sensory paradigm, and the experimental process in (c) motor paradigm and (d) sensory paradigm.
  • ...and 2 more figures