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

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Intracortical brain-machine interfaces require low-latency, energy-efficient neural decoding solutions.
• Spiking Neural Networks (SNNs) on neuromorphic hardware offer efficiency via sparse activations but face computational cost challenges for implants.

Purpose of the Study:

• To introduce a novel adaptive pruning algorithm for SNNs tailored for high-sparsity, intracortical neural decoding.
• To reduce the computational cost of SNNs for ultra-efficient neural implants.

Main Methods:

• Developed an adaptive pruning algorithm with dynamic decision adjustment and a rollback mechanism.
• Targeted SNNs with high activation sparsity for intracortical neural decoding applications.
• Evaluated performance on the NeuroBench Non-Human Primate (NHP) Motor Prediction benchmark.

Main Results:

• The pruned SNN achieved comparable decoding accuracy to dense networks.
• Demonstrated a maximum tenfold improvement in efficiency.
• Hardware simulation showed sub-μW power levels on a neuromorphic processor.

Conclusions:

• The adaptive pruning algorithm significantly enhances SNN efficiency for neural decoding.
• The approach holds promise for developing energy-constrained, ultra-efficient intracortical brain-machine interfaces.
• Enables low-overhead on-device intelligence for neural implants.