Spiking Neural Network
A spiking neural network (SNN) is a neural network whose units communicate through discrete events in time — spikes — rather than through continuous numbers passed on every step. Each neuron keeps an internal state, usually a membrane potential, that accumulates incoming input and decays over time; when the potential crosses a threshold the neuron emits a spike, resets, and goes quiet for a while. Information is carried by when spikes occur and how often, which brings time into the computation itself. In Wolfgang Maass's 1997 paper "Networks of spiking neurons: the third generation of neural network models," SNNs were framed as the third generation after threshold-gate networks and sigmoidal networks, and shown to be at least as computationally powerful as those earlier models.
The central design choice is the neuron model, which trades biological detail against cost. The Hodgkin–Huxley equations (1952) describe ion-channel dynamics in detail but are expensive to simulate. The leaky integrate-and-fire (LIF) neuron, whose origins go back to Louis Lapicque in 1907, reduces the cell to a leaky capacitor with a threshold and is by far the most common choice in practice. Izhikevich's 2003 model reproduces many firing patterns with just two variables, offering a middle ground. Synapses can add their own dynamics, delays and plasticity rules such as spike-timing-dependent plasticity (STDP), in which a synapse strengthens when the input spike precedes the output spike and weakens when it follows. A deep learning "neuron" by contrast has no memory of its own between inputs; the spiking neuron's state is what lets it integrate evidence over time.
The appeal of SNNs is event-driven computation: work happens only when a spike occurs. If activity is sparse, most neurons do nothing most of the time, and hardware can in principle skip the corresponding multiplications and memory traffic. This matches event cameras (dynamic vision sensors), which output asynchronous per-pixel brightness changes instead of full frames, and low-power sensing tasks such as keyword detection or gesture recognition. Dedicated neuromorphic chips exploit it: IBM's TrueNorth, described in a 2014 Science paper by Merolla and colleagues, put about one million neurons and 256 million synapses on a chip with 4,096 cores at very low power, and Intel's Loihi (2018) added on-chip learning, followed by Loihi 2 in 2021. The SpiNNaker project at the University of Manchester takes a different route, using many small processors to simulate spiking networks.
Training is the field's main obstacle. A spike is an all-or-nothing step function, so its derivative is zero almost everywhere and undefined at the threshold, which breaks ordinary backpropagation. Three families of workaround exist. Local biologically inspired rules such as STDP need no gradient but have so far struggled to match deep networks on hard benchmarks. Conversion methods train a conventional network and map its activations to spike rates, which works but can require many time steps and loses the latency advantage. Surrogate-gradient methods, summarized by Neftci, Mostafa and Zenke in 2019, replace the step function's derivative with a smooth stand-in during the backward pass and unroll the network through time, which has become the dominant approach. Even so, SNNs trail transformers and convolutional networks on large-scale language and vision benchmarks, and the tooling is less mature.
The headline claim of low energy use needs care. Energy savings depend on spike sparsity, the number of time steps needed, the memory access pattern and the hardware that runs the network; an SNN simulated on a GPU receives none of the event-driven benefit, and comparisons that count only multiply operations while ignoring memory and communication can overstate the gain. Neuromorphic chips are also not general replacements for GPUs: they shine on small, always-on, sensor-driven workloads and remain research platforms for most others. A further misconception is that SNNs are brain simulations — they borrow spikes and time, but typical models are far simpler than real neurons, and no one claims that a standard LIF network is how the brain computes.
Read spiking neural network as a bet that putting time and sparse events into the neuron itself can buy efficiency and temporal sensitivity, a bet that pays off on event-driven sensing and neuromorphic hardware and is still constrained by how hard it is to train.