SpikeQ: Spike-Guided Post-Training Quantization for Spiking Neural Networks

What it claims

SpikeQ is a post-training quantization (PTQ) method for spiking neural networks (SNNs). It rests on one finding: accuracy loss tracks the fraction of spikes an error flips, not the error's mean squared error (MSE). On a full-precision PreAct-ResNet18 on CIFAR-100 (74.03% top-1), injected errors of equal MSE gave very different accuracy. Constant input-current errors cut accuracy to about 19.3%, while zero-sum errors of the same MSE did far less damage. Membrane-potential errors that crossed the threshold θ dropped accuracy to about 47.6%; sign-flipped errors that never crossed θ caused almost no drop. The Spearman correlation between accuracy and flipped spikes is −0.99, against −0.77 for MSE. Existing SNN PTQ methods (AdaRound, GPTQ, S-OPTQ, SBC) all minimize numerical error.

How it works

Weights. Each layer is quantized with a spike-based objective:

• Count term: spike-count error against two references, mixed by ρ ∈ [0,1]. One is the full-precision network. The other is the full-precision layer weights applied to the already-quantized input spikes.
• Margin term: sums |V̂ − θ| over wrong spikes. It breaks ties when one level change moves no spike.
• Population term: keeps the layer's total spike count on target. Its weight μ is adjusted after each sweep and kept below 1, so it never outweighs fixing a neuron's spike.

There is no gradient through spikes, so search is discrete. Each weight picks from its four nearest levels, not two.

flowchart TD
  A[Phase 1: inputs by largest current] --> B[Pick level from 4 nearest, undecided weights as random noise]
  B --> C[Phase 2 sweep: keep ±1 changes that lower loss]
  C --> D{Held-out loss rose?}
  D -->|yes| E[Restore and freeze channel]
  D -->|no| F[Update μ]
  E --> F
  F --> C

Membrane potentials. Most prior methods quantize the post-reset state. Neuromorphic chips such as Loihi and Speck instead store V at a fixed bit width and compare the stored value with θ. SpikeQ targets this pre-fire case.

Its grid places θ midway between two levels, so rounding never flips the current-step spike for any step size. The grid also keeps the post-reset state on a level:

• Hard reset: 0 is a level, with one level above θ. The step is Δv = θ/(m+½).
• Soft reset: θ is a whole number of steps. The step is Δv = θ/m, and the number of levels below θ is calibrated.

Per layer, all (m, n) candidates are tried. The grid with the fewest spike mismatches against full-precision membrane simulation wins.

Evidence

The abstract reports results on seven benchmarks covering classification, segmentation and detection, all at T = 4 time steps. With 2-bit pre-fire membrane potentials, every existing grid collapses, while SpikeQ keeps most full-precision accuracy. At 4-bit weights and potentials, accuracy stays within 2.59 points of full precision, with 7.8 to 7.9 times less memory and 6.4 to 6.9 times less energy.

The visible models are:

• PreAct-ResNet-18 (soft reset) on CIFAR-10 and CIFAR-100.
• SEW-ResNet-18 and Spike-driven Transformer 8-384 on ImageNet.
• 55M SDT-V2 for segmentation.

Limits

The fetched text stops at the experimental setup, so per-benchmark tables, baselines, the detection setup and the 2-bit numbers are not visible. The calibration set size, ρ and λ values, and search runtime are not given. All stated experiments use T = 4. The energy figures' measurement method is not visible. The page names no code release or licence.