A Workload-Aware Energy Comparison of STDP and Surrogate-Gradient Spiking Neural Networks on MNIST

Spiking neural networks (SNNs) are often proposed as an energy-efficient alternative to conventional neural networks due to sparse, event-driven computation and compatibility with neuromorphic hardware. However, comparisons between biologically inspired local learning (e.g., STDP) and supervised gradient-based training (via surrogate gradients) are frequently difficult to interpret: training and inference costs are often conflated, absolute energy numbers depend strongly on hardware, and conclusions can change with deployment workload.We present a controlled comparison of two MNIST digit-recognition pipelines: (i) an unsupervised STDP-based SNN implemented in Brian2 with competitive excitatory--inhibitory dynamics and homeostatic regulation, and (ii) a supervised surrogate-gradient SNN baseline implemented in PyTorch. We instrument both pipelines to log reproducible activity statistics (spike counts by population and synaptic-event proxies) and translate these into model-based energy estimates under explicitly stated assumptions.Crucially, we separate training cost from inference cost and adopt a lifetime framing: we compute breakeven-style conditions and a decision framework indicating when a training-heavy method becomes worthwhile once amortised across downstream inference volume. The result is an interpretable account of efficiency--performance trade-offs across deployment scenarios, rather than a single opaque "winner".