Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Adaptively Pruned Spiking Neural Networks for Energy-Efficient Intracortical Neural Decoding

Francesca Rivelli, Martin Popov, Charalampos S. Kouzinopoulos, Guangzhi Tang

TL;DR

Intracortical neural decoding must operate under strict power constraints. The authors propose an adaptive, per-layer pruning method tailored for high-activation-sparsity SNNs to reduce synaptic operations while preserving accuracy. On NeuroBench NHP Motor Prediction data, pruned SNNs match dense performance with up to 10× fewer effective synaptic operations, and hardware simulations on SENECA show sub-$\mu$W power. This work advances energy-efficient on-device neural decoding with potential for integration with quantization and extension to recurrent SNNs for implantable BMIs.

Abstract

Intracortical brain-machine interfaces demand low-latency, energy-efficient solutions for neural decoding. Spiking Neural Networks (SNNs) deployed on neuromorphic hardware have demonstrated remarkable efficiency in neural decoding by leveraging sparse binary activations and efficient spatiotemporal processing. However, reducing the computational cost of SNNs remains a critical challenge for developing ultra-efficient intracortical neural implants. In this work, we introduce a novel adaptive pruning algorithm specifically designed for SNNs with high activation sparsity, targeting intracortical neural decoding. Our method dynamically adjusts pruning decisions and employs a rollback mechanism to selectively eliminate redundant synaptic connections without compromising decoding accuracy. Experimental evaluation on the NeuroBench Non-Human Primate (NHP) Motor Prediction benchmark shows that our pruned network achieves performance comparable to dense networks, with a maximum tenfold improvement in efficiency. Moreover, hardware simulation on the neuromorphic processor reveals that the pruned network operates at sub-$μ$W power levels, underscoring its potential for energy-constrained neural implants. These results underscore the promise of our approach for advancing energy-efficient intracortical brain-machine interfaces with low-overhead on-device intelligence.

Adaptively Pruned Spiking Neural Networks for Energy-Efficient Intracortical Neural Decoding

TL;DR

Intracortical neural decoding must operate under strict power constraints. The authors propose an adaptive, per-layer pruning method tailored for high-activation-sparsity SNNs to reduce synaptic operations while preserving accuracy. On NeuroBench NHP Motor Prediction data, pruned SNNs match dense performance with up to 10× fewer effective synaptic operations, and hardware simulations on SENECA show sub-W power. This work advances energy-efficient on-device neural decoding with potential for integration with quantization and extension to recurrent SNNs for implantable BMIs.

Abstract

Intracortical brain-machine interfaces demand low-latency, energy-efficient solutions for neural decoding. Spiking Neural Networks (SNNs) deployed on neuromorphic hardware have demonstrated remarkable efficiency in neural decoding by leveraging sparse binary activations and efficient spatiotemporal processing. However, reducing the computational cost of SNNs remains a critical challenge for developing ultra-efficient intracortical neural implants. In this work, we introduce a novel adaptive pruning algorithm specifically designed for SNNs with high activation sparsity, targeting intracortical neural decoding. Our method dynamically adjusts pruning decisions and employs a rollback mechanism to selectively eliminate redundant synaptic connections without compromising decoding accuracy. Experimental evaluation on the NeuroBench Non-Human Primate (NHP) Motor Prediction benchmark shows that our pruned network achieves performance comparable to dense networks, with a maximum tenfold improvement in efficiency. Moreover, hardware simulation on the neuromorphic processor reveals that the pruned network operates at sub-W power levels, underscoring its potential for energy-constrained neural implants. These results underscore the promise of our approach for advancing energy-efficient intracortical brain-machine interfaces with low-overhead on-device intelligence.

Paper Structure

This paper contains 15 sections, 3 equations, 4 figures, 5 tables, 1 algorithm.

Table of Contents

  1. INTRODUCTION
  2. Method
  3. Spiking Neural Network
  4. Leaky Integrate-and-Fire (LIF) Neurons
  5. Membrane potential update
  6. Spiking and reset
  7. Adaptive Pruning for Spiking Neural Network
  8. Experiments and Results
  9. Datasets and Experiments
  10. NeuroBench Harness and Selected Metrics
  11. Comparisons with Baseline Dense Networks
  12. Comparing Global and Per-layer Pruning
  13. Ablation Studies on Adaptive Pruning
  14. Hardware Simulation on Neuromorphic Processor
  15. Conclusion and Discussion

Figures (4)

  • Figure 1: Spiking neural network with 3 hidden layers for decoding recorded neural signals of the Primate Reaching task. The input, hidden, and output dimensions are listed. The adaptive pruning algorithm is applied to the hidden layers of the network.
  • Figure 2: General steps of our proposed adaptive pruning algorithm for SNN. The pruning process keeps iterating until either the target pruned connection percentage (pruned) is reached or the adaptive pruning rate reaches the minimal value (pr).
  • Figure 3: Training and validation loss during the adaptive pruning process using the per-layer approach.
  • Figure 4: Training and validation loss during the adaptive pruning process using the global approach.