Neuromorphic Brain-Computer Interface: Research on Energy-Efficient Decoding of Sparse Spiking Neural Networks

This article focuses on the field of neuromorphic brain-computer interfaces (BCIs), exploring the energy efficiency optimization of sparse event-driven LIF spiking neural networks (SNNs) in motor cortex velocity decoding. Core research questions include: 1) The impact of sparse spike events on decoding performance; 2) Can SNNs achieve ultra-low energy consumption while matching the accuracy of strong linear decoders? Based on the NLB MC_RTT benchmark dataset, key conclusions are: Temporal context (approximately 200ms history) dominates decoding performance; SNNs are comparable to linear baselines in accuracy but significantly more energy-efficient; retaining more than 25% of spike events can maintain effective decoding.

Implantable BCIs face strict power consumption and bandwidth constraints, requiring real-time transmission and decoding of neural signals under low power. Traditional methods consume high energy when processing dense spike streams. This study aims to answer: How much motor cortex velocity decoding accuracy can be maintained when only part of the spike events are retained? Can event-driven SNNs provide a feasible path for low-power implant hardware while matching the accuracy of linear decoders?

Dataset and Preprocessing

Using the NLB MC_RTT dataset (DANDI dandiset 000129) containing primary motor cortex spike data, 50ms binning generates count matrices and sparse event lists, with the target being 2D cursor velocity.

Data Segmentation

Time-continuous segmentation (70/15/15 training/validation/testing) with gaps at boundaries to prevent label leakage.

Event Budget

Retain the earliest max(1, ⌊f·n⌋) spikes per time window, reconstruct counts to ensure consistent decoder input.

Decoder Comparison

• Ridge Regression (counts): L2-regularized linear mapping
• Ridge Regression + History: Add lag features from the previous 4 windows (200ms)
• Trained SNN (BPTT): LIF hidden layer + surrogate gradient training
• Reservoir SNN: Fixed random projection + ridge regression readout

Control Experiments

Include null hypothesis tests such as sequence shuffling, phase randomization, neuron permutation, and cyclic shifting.

Key Role of Temporal Context

Memoryless decoders (current window only) have R²≈0.17; adding 200ms history increases it to 0.5-0.54, indicating motor information is encoded in the temporal evolution pattern of firing rates.

SNN vs. Linear Baselines

Trained LIF SNNs achieve R²=0.54 on full data, comparable to ridge regression with history (0.51); accuracy does not exceed but energy efficiency is better.

Signal-Bearing Factors

Decoding depends on window firing rates and temporal alignment with movement: Shuffling spike sequences, randomizing timing, or permuting neuron identities have little effect on R², but disrupting alignment causes performance collapse.

Cost of Sparsification

When retention ratio is below 25%, decoding performance drops to random levels.

The core advantage of SNNs lies in energy efficiency:

• Event-driven computing: Synaptic operations are only performed when spikes occur, avoiding dense matrix multiplications.
• Energy consumption comparison: SNNs require ~20 synaptic operations per prediction, costing ~46 nanojoules/prediction on Loihi2 hardware (23 picojoules/synapse).
• Comparison with traditional hardware: CPU ~1 nanojoule/MAC, A100 GPU ~30 picojoules/MAC, Loihi2 ~23 picojoules/synapse, NorthPole ~2 picojoules/synapse. This energy efficiency improvement is critical for implantable devices, reducing power consumption and heat dissipation.

Performance of each decoder on full data (f=1.0):

• Ridge Regression (counts only): R²=0.168±0.013
• Ridge Regression + 4-window history: R²=0.509±0.020
• Trained SNN (BPTT): R²=0.542±0.021

As sparsification increases (f=0.50/0.25/0.10), performance of all decoders decreases, but those with history features remain superior. At f=0.10, performance drops to random levels.

Comparison with NLB MC_RTT leaderboard: The results of this study (0.51-0.54) are comparable to strong linear/GPFA/SLDS baselines (0.49-0.58), while Transformer (NDT 0.62) and latent dynamics methods (AutoLFADS 0.67, MINT 0.69) are better.

Research Significance

1. Temporal context priority: Neural data can be compressed to reduce transmission bandwidth without significant performance loss.
2. SNN practicality: Accuracy is comparable to traditional methods, but event-driven characteristics are suitable for power-constrained scenarios.
3. Feasibility of sparse coding: Retaining 25-50% of spike events is still acceptable, providing space for energy efficiency optimization.

Future Directions

Explore more efficient SNN architectures, adaptive sparsification strategies, and deployment validation on actual neuromorphic hardware.