ANN-to-SDNN Conversion via Group Neurons

Updated 13 February 2026

The paper introduces a conversion method where replacing standard IF neurons with Group Neurons reduces quantization error, enabling near-lossless performance within two to four time steps.
The methodology reuses pre-trained ANN weights by substituting neurons with groups that use staggered thresholds, eliminating the need for retraining while optimizing latency-accuracy trade-offs.
Empirical benchmarks on CIFAR-10/100 and ImageNet show that the Group Neuron approach achieves performance within 0.1% of the source ANN while significantly lowering mean squared error.

A conversion from Artificial Neural Networks (ANNs) to Spiking Deep Neural Networks (SDNNs) enables leveraging energy-efficient, event-driven computation of spiking systems while reusing the advances of ANN training. This process, known as ANN-to-SDNN (or more specifically, ANN-to-SNN) conversion, is a cornerstone methodology for deploying high-performance SNNs, particularly when direct SNN training is costly or ineffective. This article synthesizes the state-of-the-art in ANN-to-SDNN conversion, focusing on theoretical motivations, algorithmic formulations, error analysis, practical workflows, and benchmark results, with particular emphasis on innovations such as Group Neurons for improved accuracy-latency trade-offs.

1. Theoretical Foundations of ANN-to-SDNN Conversion

Traditional ANN-to-SNN conversion builds on the observation that rate-coded SNNs can approximate ReLU–activated ANNs by matching the firing rate of each integrate-and-fire (IF) neuron to the normalized activation of its ANN counterpart over a finite simulation window of $T$ time steps. With standard reset-by-subtraction IF dynamics and threshold balancing (i.e., setting the SNN threshold $\theta^l$ equal to the clipping/quantization limit $\lambda^l$ of the source ANN layer), the SNN converges to the correct activation as $T \rightarrow \infty$ , but is subject to quantization error at low $T$ : $\Delta_{\rm IF} = \frac{\theta}{T}, \qquad \epsilon_{\max}^{\rm IF} = \frac{\theta}{2T}$ This quantization manifests as a trade-off between inference latency (governed by $T$ ) and output fidelity. The optimal conversion framework formalizes these relations, showing exact match in the ideal limit and giving layer-wise fit bounds and error terms (Ding et al., 2021).

2. Group Neurons: Reducing Quantization Error

To mitigate the limitations imposed by coarse quantization at low $T$ , the Group Neuron (GN) construct replaces each standard IF neuron with a group of $K$ lateral-inhibitory IF units, each with a staggered threshold:

Members $k=1, \dots, K$ possess thresholds $\theta_k = \frac{k}{K}\theta$ .
Upon firing, all group members are “soft-reset” by the group increment $\theta_{\mathrm{GN}} = \frac{\theta}{K}$ .
The group’s output at each step is $o_t = \sum_{k=1}^K o_t^{(k)}$ .

The quantization step thus becomes $\Delta_{\rm GN} = \frac{\theta}{K T}$ , yielding up to $K$ -fold reduction in maximum approximation error: $\epsilon_{\max}^{\rm GN} = \frac{1}{K}\epsilon_{\max}^{\rm IF}$ Consequently, GNs enable much finer rate approximation for a given $T$ , dramatically lowering the conversion gap—making near-lossless ANN-to-SNN conversion feasible within two to four time steps (see §4 below) (Lv et al., 2024).

3. Conversion Workflow with Group Neurons

The practical workflow for conversion under the GN framework is as follows:

Begin with a trained ReLU/QCFS-based ANN, recording per-layer clipping/activation thresholds $\lambda^l$ .
Replace each IF neuron (threshold $\theta^l$ $θ^{l}$ ) in the canonical SNN conversion with a group of $K$ $K$ IF members per neuron:
- For each group member, set $\theta_k^l = \frac{k}{K} \theta^l$ .
- Use the same layer weights $W^l$ for all group members; no retraining or reweighting is required.
- Configure the group reset increment $\theta^{l}_{\mathrm{GN}} = \theta^l / K$ for the lateral-inhibitory dynamics.
No architectural, weight, or training modifications to the underlying ANN are needed. All enhancements occur strictly post hoc by substituting neuron microcircuits (Lv et al., 2024).

4. Quantitative Performance and Latency-Accuracy Trade-off

Empirical benchmarks on CIFAR-10/100 and ImageNet demonstrate the efficacy of GNs:

GN-based SNNs reach accuracy within $<0.1\%$ of the source ANN with $T = 2$ –$4$.
For ResNet-18 on CIFAR-10 (ANN $=96.48\%$ $= 96.48%$ ), SNN-GN achieves
- $T=1$ : $96.01\%$ , $T=2$ : $96.36\%$ , $T=4$ : $96.57\%$ .
On ImageNet (ResNet-34, ANN $=74.35\%$ ): $T=2$ : $73.61\%$ , $T=4$ : $73.73\%$ .
GN reduces the mean squared error between final SNN rates and ANN activations by roughly $10\times$ compared to a standard IF implementation at low $T$ .
The method consistently matches or outperforms other state-of-the-art conversions (e.g., QCFS, OPI, COS, SRP) on accuracy and low-latency efficiency (Lv et al., 2024).

$T$	SNN–IF MSE	SNN–GN MSE	CIFAR-100 Acc (VGG-16, $T=2$ )	ImageNet Acc (ResNet-34, $T=4$ )
1	11.65	1.39	—	—
2	6.36	0.64	76.36%	—
4	3.72	0.37	—	73.73%

Experimentally, accuracy “flattens out” after $T=2$ –$4$, in contrast to vanilla IF, which requires $T\gg10$ for similar fidelity.

5. Error Analysis and Rate-Approximation Theory

Summing over $T$ , the GN output’s mean coincides with a high-resolution approximation of the ANN activation with reduced quantization: $\Phi_{\mathrm{GN}}(T) = W\,\Phi_{\mathrm{GN}}^{\text{prev}}(T) - \frac{u_{T+1} - u_{1}}{T}$ Residual error thus decomposes as a rate-proximate output plus a vanishing $O(1/T)$ term, with quantization steps $1/(K T)$. GNs provide not only a higher granularity per time step but also exploit the superposition of multiple staggered-threshold accumulators (Lv et al., 2024).

6. Comparison to Other Advanced Conversion Techniques

The GN approach subsumes or complements several other conversion innovations:

Noise-aware conversion incorporates residual error modeling at training time but leaves SNN dynamics unchanged (Huang et al., 2024).
Adaptive thresholding and confidence gating offer orthogonal improvements for highly dynamic or non-stationary input statistics (He et al., 2023).
Hybridization, burst coding, and multi-threshold extension (as in Transformers/ViTs) provide alternative means of reducing quantization and energy cost per inference, but do so through structurally or functionally different mechanisms (Huang et al., 28 Feb 2025, Wang et al., 2024).
GNs particularly excel in ultra-low-latency and hardware-aligned contexts, since they avoid architectural or training complexity increases (Lv et al., 2024).

7. Practical Considerations and Implementation

GN-based conversion is weight- and topology-agnostic and is compatible with the majority of modern ANN architectures and their pre-trained models. It can be integrated into open-source pipelines (code available at https://github.com/Lyu6PosHao/ANN2SNN_GN). The only hyperparameter required is the group size $K$ ( $K=4$ for standard settings, up to 6 for large-scale ImageNet architectures). Empirical studies confirm that this minimal additional neuron-level complexity yields significant accuracy and efficiency gains without retraining, offering state-of-the-art trade-offs for both academic benchmarks and neuromorphic deployment scenarios (Lv et al., 2024).

By integrating finely quantized multi-member dynamics at the neuron level, ANN-to-SDNN conversion via Group Neurons enables near-exact and ultra-fast conversion, setting a new empirical and theoretical standard for efficient spiking deep networks deployable at low latency and low energy budgets.