Research map / Architectures and efficiency / Spiking networks

Exploring the Limitations of Layer Synchronization in Spiking Neural Networks

NeurIPS 2026 · Architectures and efficiency › Spiking networks · approach: Training deep spiking networks · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows that SNNs trained with per-layer synchronization degrade or lose efficiency under asynchronous neuron execution, and proposes "unlayered backpropagation" that trains through a scheduled asynchronous forward pass (random and momentum scheduling), recovering accuracy with fewer spikes and lower latency.

Problem

SNNs are trained with layer-synchronized backpropagation, but neuromorphic hardware that runs asynchronously (event-driven, no per-layer barriers) then shows a mismatch in dynamics, accuracy loss, and little energy or latency gain.

Approach

Simulate asynchronous, per-spike network execution with a neuron scheduling policy (random or momentum) inside the forward pass and backpropagate through that unrolled schedule ("unlayered backpropagation"), so weights are learned for asynchronous dynamics; forward group size F interpolates between layered and fully asynchronous execution.

Key results

• On N-MNIST, layered-trained models drop from 0.949 (layered inference) to 0.625 under async RS inference, while unlayered MS training reaches 0.963 accuracy with async MS inference.
• Unlayered MS training reaches 0.856 on DVS gesture versus 0.739 for layered training, with activation density roughly halved on N-MNIST (1.476 vs 3.987).
• Projected energy on µBrain falls from 0.9155 µJ (layered) to 0.3389 µJ (unlayered MS, async) on N-MNIST, with synaptic operations reduced by about 2.7x.

Surprising findings

• Asynchronous models are sparser when inference stops at the first output decision, even though they can have more total spikes when the network is fully drained.
• Layered-trained models lose much accuracy under asynchronous inference (N-MNIST 0.949 to 0.625), even though the asynchronous mode is supposed to be a natural fit for neuromorphic hardware.

Limitations

• Unlayered backpropagation is computationally expensive, with time and memory roughly doubling each time the forward group size F is halved.
• Energy figures are projections from energy per synaptic operation reported in the literature rather than on-hardware measurements.
• Experiments use narrow, deep fully connected or small VGG-style networks, and the approach has not been tested on large modern models.

Experimental setup

Datasets: N-MNIST SHD DVS gesture CIFAR-10

Benchmarks: N-MNIST SHD DVS gesture CIFAR-10

Metrics: accuracy activation density synaptic operations energy (µJ) forward steps (latency)

Baselines: Layered training with layered inference Layered training with async RS inference

Builds on: surrogate gradient backpropagation through time leaky integrate-and-fire neuron EventProp SparseProp

Closest papers by meaning

• From Cortical Synchronous Rhythm to Brain Inspired Learning Mechanism: An Oscillatory Spiking Neural Network with Time-Delayed Coordination (NeurIPS 2026 · similarity 0.88)
• Spiking Token Mixer: An event-driven friendly Former structure for spiking neural networks (NeurIPS 2024 · similarity 0.87)
• Temporal Flexibility in Spiking Neural Networks: Towards Generalization Across Time Steps and Deployment Friendliness (ICLR 2025 · similarity 0.87)
• Rethinking Spiking Neural Networks from an Ensemble Learning Perspective (ICLR 2025 · similarity 0.86)
• Learning Delays in Spiking Neural Networks using Dilated Convolutions with Learnable Spacings (ICLR 2024 · similarity 0.86)
• A Scalable, Causal, and Energy Efficient Framework for Neural Decoding with Spiking Neural Networks (NeurIPS 2025 · similarity 0.86)