Compute Accounting in Hybrid ANN-SNN Architectures on Fashion-MNIST

Hybrid ANN-SNN architectures are often presentedas more efficient than pure spiking networks because they fire farfewer spikes. We show that this framing is misleading, becauseit hides the dense computation the hybrid still performs inits non-spiking layers. We train three matched neural networkarchitectures on Fashion-MNIST and compare them on accuracyand per-image compute: a standard convolutional ANN, a fullyspiking SNN with leaky integrate-and-fire neurons, and a hybridmodel with an ANN convolutional front-end feeding a spikingclassifier head. All three architectures share identical layer shapesand exactly 1,199,882 parameters, and are trained under the samehyperparameters. Across five random seeds the three modelsreach similar clean test accuracy (ANN 92.82 ± 0.25%, Hybrid92.66 ± 0.25%, SNN 92.00 ± 0.16%), but their compute profilesdiffer sharply. The hybrid model fires 2,602 synaptic operationsper image in its spiking classifier head, compared to 13.5 millionfor the fully spiking network. Reported alone, this looks like a5,000-fold efficiency gain for the hybrid. It is not. The hybrid stillperforms roughly 12 million multiply-accumulate operations perimage in its non-spiking convolutional layers, which are invisibleto a SynOps-based metric. A fair comparison reports both kindsof operations separately and does not combine them without anexplicit hardware energy model. This accounting choice changesthe apparent efficiency of hybrid spiking architectures by ordersof magnitude.