Polarity Multi-Spike Mapping (PMSM)

• PMSM is an ANN-to-SNN conversion method that couples an entropy-guided quantizer with polarity-aware neurons to reconstruct activations nearly losslessly in a single timestep.
• It introduces Polarity Quantized Activation (PQA) to preserve both positive and negative information, addressing representational loss post-Batch Normalization.
• Empirical evaluations show state-of-the-art accuracy parity with ANNs and over 5× energy reduction on benchmarks like CIFAR and ImageNet.

Searching arXiv for the specified paper and closely related conversion work to ground the article. Polarity Multi-Spike Mapping (PMSM) is an ANN-to-SNN conversion method introduced in "Quantization Meets Spikes: Lossless Conversion in the First Timestep via Polarity Multi-Spike Mapping" that couples an information-theoretic quantizer with a polarity-aware spiking neuron model to achieve nearly lossless conversion at the first timestep while preserving stable performance over multiple timesteps (Zhang et al., 20 Aug 2025). The method addresses two persistent conversion bottlenecks: the mismatch between continuous ANN activations and discrete spike trains, and the loss of negative-value information in conventional quantized activation functions after Batch Normalization (BN). Its core components are Polarity Quantized Activation (PQA), Augmented Integrate-and-Fire (AIF) neurons, an entropy-guided threshold strategy, and a conversion rule under which the reconstruction error at $T=1$ is approximately zero.

1. Problem setting and motivation

PMSM is situated within conversion-based spiking inference rather than direct SNN training. The paper motivates this choice by noting that direct training methods address the non-differentiability of SNN activations but often incur high computational and energy costs during training, whereas ANN-to-SNN conversion remains a valuable and practical alternative (Zhang et al., 20 Aug 2025).

The method targets two fundamental gaps. First, ANN activations are continuous and real-valued, whereas SNN outputs are discrete spikes; this produces layer-wise accumulation of approximation errors during conversion. Second, conventional quantized activation functions discard negative information after BN, which reduces information capacity. In the PMSM formulation, these issues are not treated as separate engineering artifacts but as linked sources of representational loss.

The paper’s proposed resolution is tripartite. It designs an entropy-maximizing quantized activation, PQA, that preserves both positive and negative information; it introduces an AIF neuron that emits multiple polarity-aware spikes in a single timestep; and it proves that, under specific threshold and initialization conditions, the conversion error at $T=1$ is nearly zero. This framing is important because the theoretical minimum latency for converted SNNs is one timestep, yet the paper states that existing conversion methods have struggled to realize such ultra-low latency without accuracy loss.

2. Information-entropy formulation and Polarity Quantized Activation

The theoretical analysis begins with the instantaneous pre-activation $x \sim \mathcal{N}(0,1)$, interpreted as the output of a folded BN layer. Its differential entropy is given as

$\mathcal{H}_{BN} = \frac{1}{2}\log(2\pi e).$

A conventional ReLU-based quantized activation with threshold $\theta$ and $L$ levels is written as

$$f_{QA}(x)= \theta\cdot \mathrm{clip}\!\left(\lfloor xL/\theta \rfloor/L,\,0,\,1\right).$$

Under this construction, all negative values are removed, and the maximum entropy is reported as

$$\mathcal{H}_{QA}^{\max} = \mathcal{H}_{ReLU} \approx 0.69\,\mathcal{H}_{BN},$$

which the paper describes as a $31\%$ drop in representational capacity (Zhang et al., 20 Aug 2025).

PMSM replaces this with Polarity Quantized Activation. For layer $l$, the PQA function is defined as

$T=1$0

with

$T=1$1

The clipping interval $T=1$2 allocates $T=1$3 quantization bins across negative and positive values. Unlike a ReLU-quantizer, PQA retains negative bins explicitly.

The entropy of PQA depends on the discretized mass

$T=1$4

for $T=1$5, together with clipping mass for $T=1$6 and $T=1$7. The central criterion is the entropy ratio

$T=1$8

Grid search over $T=1$9 and $x \sim \mathcal{N}(0,1)$0 for fixed $x \sim \mathcal{N}(0,1)$1 and $x \sim \mathcal{N}(0,1)$2 identifies regions where $x \sim \mathcal{N}(0,1)$3, which the paper interprets as lossless quantization. This suggests that the negative branch is not an auxiliary refinement but the main mechanism by which the quantizer recovers the entropy deficit introduced by ReLU truncation.

3. Augmented Integrate-and-Fire neurons and polarity multi-spikes

PQA alone does not complete the conversion; its quantized outputs must be reconstructed by a spiking neuron in one timestep. PMSM therefore introduces the Augmented Integrate-and-Fire neuron, whose role is to emit multiple signed spikes in a single update (Zhang et al., 20 Aug 2025).

For timestep $x \sim \mathcal{N}(0,1)$4, the neuron dynamics are

$x \sim \mathcal{N}(0,1)$5

$x \sim \mathcal{N}(0,1)$6

$x \sim \mathcal{N}(0,1)$7

$x \sim \mathcal{N}(0,1)$8

The threshold is set as

$x \sim \mathcal{N}(0,1)$9

so that $\mathcal{H}_{BN} = \frac{1}{2}\log(2\pi e).$0 and the quantizer scale matches the spiking threshold. The spike limits are

$\mathcal{H}_{BN} = \frac{1}{2}\log(2\pi e).$1

which bound the number of negative and positive spikes that can be emitted in one step. The reset is soft: the membrane subtracts exactly $\mathcal{H}_{BN} = \frac{1}{2}\log(2\pi e).$2, preserving residual potential.

The mapping algorithm is correspondingly direct. A pretrained ANN with weights and BN parameters is converted by folding BN into $\mathcal{H}_{BN} = \frac{1}{2}\log(2\pi e).$3, computing optimal $\mathcal{H}_{BN} = \frac{1}{2}\log(2\pi e).$4 via entropy-guided tuning, setting $\mathcal{H}_{BN} = \frac{1}{2}\log(2\pi e).$5, replacing ReLU+BN+quantizer with PQA and AIF neurons, initializing each membrane with $\mathcal{H}_{BN} = \frac{1}{2}\log(2\pi e).$6, and running the network from $\mathcal{H}_{BN} = \frac{1}{2}\log(2\pi e).$7. The algorithm also permits inference with $\mathcal{H}_{BN} = \frac{1}{2}\log(2\pi e).$8 timesteps by accumulating spikes through the stated temporal equation.

4. Error analysis and hyperparameter strategy

The conversion analysis is expressed through the average post-synaptic potential over $\mathcal{H}_{BN} = \frac{1}{2}\log(2\pi e).$9 steps:

$\theta$0

The absolute error relative to the ANN activation $\theta$1 is

$\theta$2

Under static input $\theta$3 and proper initialization $\theta$4, the paper states two asymptotic regimes (Zhang et al., 20 Aug 2025). At $\theta$5, $\theta$6 exactly, yielding perfect reconstruction of $\theta$7 in a single step. For $\theta$8,

$\theta$9

so temporal averaging drives $L$0 as $L$1.

The hyperparameter adjustment strategy is aligned with the entropy analysis. The procedure fixes $L$2 and searches $L$3 to maximize $L$4. The paper reports that the empirical optimum lies near $L$5, equivalently $L$6, with $L$7 and $L$8 for CIFAR. Figure 1 is described as exhibiting a white region where $L$9, and this region guides hyperparameter selection.

A common simplification in ANN-to-SNN conversion is that one-timestep inference necessarily incurs unrecoverable discretization error. PMSM explicitly contests that view, but only under the stated conditions of threshold matching, signed multi-spike firing, and membrane initialization. The paper’s exactness claim is therefore conditional rather than unconditional.

5. Reported empirical performance and energy characteristics

The reported experimental results span CNNs and ViTs on CIFAR-10, CIFAR-100, and ImageNet (Zhang et al., 20 Aug 2025). On VGG-16 and ResNet-20, the paper reports exact parity between ANN and SNN at one timestep. On ViT-S, the one-step SNN remains close to the ANN and modestly improves again at $$f_{QA}(x)= \theta\cdot \mathrm{clip}\!\left(\lfloor xL/\theta \rfloor/L,\,0,\,1\right).$$0.

Setting	Reported accuracies	Note
VGG-16 on CIFAR-10/100	ANN: 95.67%; SNN@T=1: 95.67%	Table 1
ResNet-20 on CIFAR-10/100	ANN: 93.78%; SNN@T=1: 93.78%	Table 1
ViT-S on CIFAR-10	ANN 98.8%; SNN@1 step 98.5%; SNN@16 steps 98.6%	Table 2
ViT-S on CIFAR-100	ANN 90.4%; SNN@1 step 89.3%; SNN@16 steps 89.4%	Table 2
ViT-S on ImageNet	ANN 82.3%; SNN@1 step 81.6%; SNN@16 steps 81.7%	Table 2

The abstract characterizes the ViT-S one-timestep results as state-of-the-art accuracies of $$f_{QA}(x)= \theta\cdot \mathrm{clip}\!\left(\lfloor xL/\theta \rfloor/L,\,0,\,1\right).$$1 on CIFAR-10, $$f_{QA}(x)= \theta\cdot \mathrm{clip}\!\left(\lfloor xL/\theta \rfloor/L,\,0,\,1\right).$$2 on CIFAR-100, and $$f_{QA}(x)= \theta\cdot \mathrm{clip}\!\left(\lfloor xL/\theta \rfloor/L,\,0,\,1\right).$$3 on ImageNet, all with one timestep, and describes them as establishing a new benchmark for efficient conversion.

Energy is reported using a spike-count model with $$f_{QA}(x)= \theta\cdot \mathrm{clip}\!\left(\lfloor xL/\theta \rfloor/L,\,0,\,1\right).$$4 and $$f_{QA}(x)= \theta\cdot \mathrm{clip}\!\left(\lfloor xL/\theta \rfloor/L,\,0,\,1\right).$$5 on VGG-16/CIFAR. The comparison is against QCFS.

Setting	Total spikes $$f_{QA}(x)= \theta\cdot \mathrm{clip}\!\left(\lfloor xL/\theta \rfloor/L,\,0,\,1\right).$$6	Power / Ratio
QCFS, $$f_{QA}(x)= \theta\cdot \mathrm{clip}\!\left(\lfloor xL/\theta \rfloor/L,\,0,\,1\right).$$7	3.08	$$f_{QA}(x)= \theta\cdot \mathrm{clip}\!\left(\lfloor xL/\theta \rfloor/L,\,0,\,1\right).$$8 / —
PMSM, $$f_{QA}(x)= \theta\cdot \mathrm{clip}\!\left(\lfloor xL/\theta \rfloor/L,\,0,\,1\right).$$9	0.61	$$\mathcal{H}_{QA}^{\max} = \mathcal{H}_{ReLU} \approx 0.69\,\mathcal{H}_{BN},$$0 / $$\mathcal{H}_{QA}^{\max} = \mathcal{H}_{ReLU} \approx 0.69\,\mathcal{H}_{BN},$$1
QCFS, $$\mathcal{H}_{QA}^{\max} = \mathcal{H}_{ReLU} \approx 0.69\,\mathcal{H}_{BN},$$2	6.44	$$\mathcal{H}_{QA}^{\max} = \mathcal{H}_{ReLU} \approx 0.69\,\mathcal{H}_{BN},$$3 / —
PMSM, $$\mathcal{H}_{QA}^{\max} = \mathcal{H}_{ReLU} \approx 0.69\,\mathcal{H}_{BN},$$4	1.23	$$\mathcal{H}_{QA}^{\max} = \mathcal{H}_{ReLU} \approx 0.69\,\mathcal{H}_{BN},$$5 / $$\mathcal{H}_{QA}^{\max} = \mathcal{H}_{ReLU} \approx 0.69\,\mathcal{H}_{BN},$$6

These values underlie the statement that PMSM reduces energy consumption by over $$\mathcal{H}_{QA}^{\max} = \mathcal{H}_{ReLU} \approx 0.69\,\mathcal{H}_{BN},$$7 under VGG-16 on CIFAR-10 and CIFAR-100 compared to the baseline method.

6. Pipeline integration, hardware compatibility, and interpretive scope

Within the conversion pipeline, PMSM alters both training-time activation design and conversion-time neuron instantiation (Zhang et al., 20 Aug 2025). During training, ReLU+BN is replaced with $$\mathcal{H}_{QA}^{\max} = \mathcal{H}_{ReLU} \approx 0.69\,\mathcal{H}_{BN},$$8, and $$\mathcal{H}_{QA}^{\max} = \mathcal{H}_{ReLU} \approx 0.69\,\mathcal{H}_{BN},$$9 is learned via back-propagation. During conversion, BN is folded into the weights and thresholds, and AIF neurons are instantiated with thresholds $31\%$0. The paper states that the algorithmic steps require no additional fine-tuning of weights; the adaptation is concentrated in quantizer thresholds and neuron models.

The method is described as compatible with neuromorphic chips supporting multi-bit and signed spikes, including SpiNNaker and Loihi 2. This compatibility follows from the fact that PMSM does not rely on single-bit, positive-only firing. Instead, it assumes neurons can represent bounded positive and negative spike counts within one timestep.

In interpretive terms, PMSM combines two normally separate optimization targets: quantizer fidelity and spiking-time fidelity. The entropy analysis is used to justify the quantizer, while the AIF construction ensures the quantized value can be reconstructed by the SNN dynamics at $31\%$1. A plausible implication is that the method’s reported single-step performance is not merely a consequence of lower approximation error per layer, but of an explicit scale-matching relation between $31\%$2, $31\%$3, and $31\%$4.

The paper’s concluding position is that PMSM fuses an information-theoretic quantizer with a multi-spike, polarity-aware neuron to obtain near-lossless reconstruction of ANN activations in a single timestep while preserving the temporal-averaging benefits of longer inference. Within the ANN-to-SNN conversion literature, its specific contribution is therefore the claim that ultra-low-latency conversion need not sacrifice either negative activation information or temporal robustness, provided the quantization and spiking models are co-designed.