A purely discrete spiking neural network via ANN-to-SNN conversion with representation alignment and readout calibration

Abstract

Artificial neural network-to-spiking neural network (ANN-to-SNN) conversion is affected by both hidden-representation mismatch and incompatibility between an ANN-trained classifier and finite-window spike statistics. This study separates these effects using a quantization clip-floor-shift (QCFS)–Chialvo conversion model with a forced winner-take-all readout and controlled experiments. In the main stochastic QCFS–Chialvo experiments, hard-label readout recalibration fits ridge regression from validation-set firing rates to one-hot class labels. At T=400, it increased the ReLU-source Chialvo SNN accuracy from 67.198% to 92.992% on MNIST and from 39.896% to 82.483% on Fashion-MNIST. A label-free moment-alignment control that retained the ANN classifier reached 93.687% and 81.137% on MNIST and Fashion-MNIST, respectively. This result indicates that a substantial portion of the recovered accuracy is associated with distribution mismatch at the hidden-rate/readout interface, while the remaining advantage of hard-label recalibration on Fashion-MNIST suggests an additional target-dependent readout component. After hard-label recalibration, QCFS changed accuracy by only +0.003 and −0.122 percentage points, respectively. Mechanistic analysis nevertheless showed that source-side QCFS and QCFS current shaping improved global ANN–SNN representational similarity, while QCFS current shaping increased invariance to small pre-activation perturbations. Under firing-rate-matched conditions, leaky integrate-and-fire (LIF) neurons achieved higher static-image accuracy, whereas Chialvo neurons generated higher-dimensional population spike codes, stronger damped alternating temporal correlations, and greater linear memory capacity. In a separate deterministic IF comparison, teacher-logit readout recalibration fits the final linear layer to source-ANN logits without using ground-truth labels. At T=4, it improved direct conversion by 1.625 percentage points on MNIST and 3.846 points on Fashion-MNIST, reaching performance comparable to Li-style calibration; Bu-style QCFS remained the strongest ultra-low-latency method. The methodological contribution is therefore the controlled separation of representation, interface, and readout-target effects, rather than a new linear-regression solver.