Neuron Benchmark

The question

Spiking neural networks promise a more event-driven, potentially more efficient way to compute. But their behavior is determined partly by the simulated neuron inside each layer — and the tradeoff between simple models (cheap, less realistic) and biologically rich ones (slow, more realistic) is rarely measured. This benchmark makes that tradeoff visible.

Design

Instead of comparing neuron models in isolation, each one is dropped into the same network, given the same inputs, and measured on both task behavior and simulation cost:

Input (Rates) -> Encoder (Linear) -> Neuron (SNN Layer) -> Readout -> Class

Five neuron models — IF, LIF, Adaptive LIF, Izhikevich, Hodgkin-Huxley — trained with Adam + surrogate-gradient spike functions, evaluated on accuracy, spike rate, training time, and an energy proxy.

Key findings

1. LIF is the strongest general baseline — good accuracy at low cost.
2. Izhikevich offers the best accuracy-per-complexity tradeoff — richer dynamics at moderate cost.
3. Hodgkin-Huxley is the gold-standard reference — too slow for production, essential for biological validation.
4. Benchmark everything, never assume — simpler models don’t always win.

Multi-seed, not anecdote

The whole point was reproducibility: the CLI runs 5-seed, 20-epoch experiments across every model, plus a robustness pass (noise, timestep, reduced data) and an MNIST comparison. The output is plots and a report, not vibes.

python -m neuron_benchmark --experiment --neuron all --seeds 0 1 2 3 4 --full-metrics

Interactive

A Streamlit demo runs the same pipeline live so you can watch spike traces and accuracy update as you switch neuron models — which made it a much better tool for explaining the results than a static report ever would be.