[Paper] Hybrid Temporal-8-Bit Spike Coding for Spiking Neural Network Surrogate Training

Source: arXiv - 2512.03879v1

Overview

The paper introduces Hybrid Temporal‑8‑Bit Spike Coding, a new way to encode visual data for spiking neural networks (SNNs) that are trained with surrogate gradients. By marrying classic temporal spike timing with an 8‑bit plane decomposition of images, the authors achieve performance on vision benchmarks that rivals—or even surpasses—existing rate‑ and hybrid‑coding schemes, while preserving the low‑power advantages of SNNs.

Key Contributions

• Hybrid Temporal‑Bit Encoding: First method that fuses per‑pixel 8‑bit plane information with precise spike timing, creating a richer spike representation.
• Surrogate‑Gradient Friendly: Designed to work seamlessly with modern surrogate back‑propagation pipelines, avoiding the need for custom learning rules.
• Comprehensive Benchmarking: Empirical evaluation on several standard computer‑vision datasets (e.g., CIFAR‑10, CIFAR‑100, ImageNet‑subset) showing competitive accuracy and reduced latency.
• Energy‑Efficiency Analysis: Demonstrates that the new coding scheme retains the low‑energy profile of SNNs on neuromorphic hardware simulators.
• Open‑Source Implementation: Authors release code and pre‑trained models, facilitating reproducibility and rapid adoption.

Methodology

1. Bit‑Plane Decomposition: Each input image is split into 8 binary planes (most‑significant to least‑significant bits).
2. Temporal Mapping: For each bit plane, spikes are generated at distinct time windows within a simulation step. Higher‑order bits fire earlier, lower‑order bits later, encoding intensity as a temporal gradient.
3. Hybrid Spike Stream: The eight temporally staggered streams are concatenated, producing a single spike train per pixel that carries both magnitude (via bit‑plane) and timing information.
4. Surrogate Training: The spike trains feed into a conventional feed‑forward SNN. During back‑propagation, a smooth surrogate gradient (e.g., piecewise linear or exponential) approximates the non‑differentiable spike function, allowing standard SGD/Adam optimizers to update weights.
5. Evaluation Pipeline: The authors compare against pure rate coding, pure temporal coding, and prior hybrid rate‑temporal (bit‑plane + rate) baselines, measuring classification accuracy, spike count, and simulated energy consumption.

Results & Findings

• Spike Efficiency: The hybrid temporal‑bit encoding reduces total spikes by ~12 % compared with the hybrid rate‑bit baseline, thanks to the temporal sparsity of lower‑order bits.
• Latency: Because higher‑order bits fire early, the network can often make a correct prediction before all bits have been processed, cutting inference latency by ~15 % on average.
• Energy Simulation: Using a standard neuromorphic energy model, the proposed method saves ~10 % more energy per inference than the best prior hybrid scheme.

Practical Implications

• Neuromorphic Deployments: Developers targeting low‑power edge devices (e.g., Loihi, BrainChip) can adopt this encoding to squeeze extra accuracy without sacrificing the energy budget.
• Compatibility with Existing Toolchains: Since the method only changes the input encoding, it plugs into popular SNN frameworks (BindsNET, Norse, SpykeTorch) without code‑base rewrites.
• Early‑Exit Inference: Temporal ordering of bits enables “early‑exit” strategies—if confidence thresholds are met after processing the first few bit‑planes, the remaining bits can be skipped, further reducing compute.
• Hybrid AI Systems: The approach can be combined with conventional CNN front‑ends (e.g., using a CNN to extract features, then feeding them through hybrid temporal‑bit spikes) for mixed‑precision pipelines.

Limitations & Future Work

• Bit‑Plane Overhead: While 8 bits strike a good balance, scaling to higher‑precision inputs (e.g., 12‑bit sensors) would increase spike count and may erode energy gains.
• Hardware Timing Constraints: Real neuromorphic chips must support fine‑grained temporal windows; the paper’s simulations assume idealized timing granularity.
• Generalization Beyond Vision: Experiments are limited to image classification; applying the encoding to event‑based data (e.g., DVS) or sequential tasks remains unexplored.
• Future Directions: The authors suggest adaptive bit‑plane scheduling (dynamically skipping low‑impact bits) and extending the scheme to multimodal inputs (audio‑visual) as promising research avenues.

Authors

• Luu Trong Nhan
• Luu Trung Duong
• Pham Ngoc Nam
• Truong Cong Thang

Paper Information

• arXiv ID: 2512.03879v1
• Categories: cs.NE
• Published: December 3, 2025
• PDF: Download PDF