A spiking neural network (SNN) is a kind of neural network that communicates the way the brain does: with discrete electrical pulses, or spikes, fired at particular moments in time. This sets it apart from the artificial neural networks behind today’s deep learning, where each neuron outputs a smooth continuous number that flows to the next layer every step.
In a spiking model, a neuron accumulates incoming signals until it crosses a threshold, then emits a single spike and resets - and crucially, the exact timing of that spike carries information. Networks of these units stay quiet until something happens, so they can be extremely energy-efficient, and they can encode information in the patterns and timing of spikes rather than in fixed activation values. This is why spiking networks are the natural software for neuromorphic chips like TrueNorth, Loihi, and SpiNNaker.
The catch is training. The backpropagation algorithm that made deep learning work relies on smooth, differentiable activations; the abrupt, all-or-nothing spike is not differentiable, which makes spiking networks harder to train to the same accuracy. Researchers use workarounds - converting trained conventional networks into spiking ones, or approximate gradient methods - but on standard benchmarks SNNs have generally trailed ordinary deep networks.
Why business readers should care: spiking networks are the most biologically faithful branch of neural computing and the key to ultra-low-power AI hardware. They remain mostly a research direction, attractive for battery-powered sensing and edge devices where energy, not raw accuracy, is the binding constraint.