Learn · Intermediate

Spiking neural networks: computing with events and time

Spiking neural networks represent information through discrete events called spikes, with each neuron’s behavior depending on signals accumulated over time. They offer a way to model biological circuits and to design computation around sparse events. Their usefulness depends on the neuron model, timing rules, learning method, and hardware; biological inspiration alone does not make a system accurate or efficient.

A neuron that waits, leaks, and fires

An ordinary artificial neuron often takes a set of numbers, combines them, and produces another number. A spiking neuron instead maintains an internal state through time. Incoming signals change that state, and a spike occurs when a threshold is crossed.

The simplest useful analogy is a bucket with a small leak and a tipping mechanism. Incoming drops raise the water level. The leak lowers it between arrivals. When the level reaches a line, the bucket tips and sends a brief signal onward. Many small inputs arriving close together can trigger a tip, while the same inputs spread far apart may leak away first.

A leaky integrate-and-fire neuron follows this broad pattern. It integrates incoming input, loses some accumulated state over time, emits a spike at a threshold, and resets according to the model’s rules. A refractory period may temporarily prevent another spike. These are modeling choices, not a claim that real neurons are literal buckets.

Timing is part of the representation

A network can encode information in how frequently neurons fire, the timing of individual spikes, or patterns distributed across many neurons. Those alternatives are not interchangeable. A neuron that counts events over a long interval behaves differently from one that reacts to nearly simultaneous inputs.

Wolfgang Maass’s paper Networks of spiking neurons: The third generation of neural network models00011-7) established an influential theoretical treatment of computation with spiking neurons. Its importance is the recognition that timing and internal dynamics change the computational model, rather than merely giving a conventional network a biological-looking diagram.

This connects spiking systems to recurrent neural networks: both carry state forward. The distinction is that spikes introduce event-based communication and temporal dynamics with their own assumptions. A recurrent network does not become a spiking network simply because it remembers previous inputs.

Wiring does not specify a complete mind

A connectome gives a map of connections among neurons. To simulate it, researchers still need rules for how neurons evolve and how connections transmit signals. They also need inputs, initial conditions, and a definition of what observed outputs mean.

The Open Fly project illustrates the distinction. Its repository describes a published leaky integrate-and-fire fly model connected to a strategy game through a human-designed interface. The project has anatomical grounding, but the game-to-neuron and neuron-to-game translations are additional decisions. The README says the circuit did not learn to play.

A map of roads does not specify traffic speeds, driving rules, or destinations. Likewise, a connection map does not fully determine the activity or behavior of a biological animal. A faithful software implementation can reproduce its chosen equations while leaving biological fidelity unresolved.

Learning through a threshold is difficult

Spikes also complicate training with backpropagation. A hard threshold produces no gradual change over most of its range and a discontinuity where the event occurs. The ordinary derivative therefore does not provide a convenient learning signal for adjusting a weight just before a missed spike.

One approach uses a surrogate gradient: the forward simulation retains the spike rule, while the backward calculation substitutes a smoother derivative that provides a usable training signal near the threshold. Emre Neftci, Hesham Mostafa, and Friedemann Zenke explain this approach in Surrogate Gradient Learning in Spiking Neural Networks.

The substitute is a practical training device. It does not mean the hard threshold actually has that smooth derivative, or that biological brains necessarily use the same algorithm. Other approaches include local timing-dependent learning rules and conversion from conventional networks, each with different assumptions.

Sparse events do not guarantee cheap computation

A network that sends few events can avoid work if its hardware processes events only when they occur. But a simulator that updates every neuron at every tiny time step may perform substantial work even when few neurons spike. Memory access, routing, simulation duration, and the cost of the surrounding software also matter.

That is why an energy or speed claim needs a workload, hardware description, accuracy target, and measurement of the complete system. Comparing only spike counts with arithmetic operations can hide costs on both sides. The broader lesson from end-to-end speed measurement applies here too.

When evaluating a spiking demonstration, ask what the circuit receives, what its outputs control, whether anything learns, and which comparison isolates the contribution of its wiring. Those questions distinguish an interesting dynamical system from evidence of acquired intelligence. Spiking networks are a rigorous computational idea; what any particular system demonstrates depends on the experiment built around them.

Ground Truth. (2026, October 18). Spiking neural networks: computing with events and time. Ground Truth. https://groundtruth.day/learn/spiking-neural-networks.html

@misc{groundtruth:spiking-neural-networks,
  title  = {Spiking neural networks: computing with events and time},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/learn/spiking-neural-networks.html}
}

Topics: fundamentals · architectures · neuroscience · spiking-neural-networks