What is a spiking neural network (SNN)? How spike-based AI works and how it relates to real neurons

A spiking neural network (SNN) is an artificial neural network in which neurons communicate through discrete spikes in time rather than continuous numbers. Each model neuron integrates incoming signals and fires a spike when its internal state crosses a threshold. Spiking neural networks are closer to biological neurons than deep learning models and are the native software of neuromorphic hardware.

Key takeaways

• In an SNN, information is carried by when spikes happen and how often, not by real-valued activations.
• Maass (1997) described spiking networks as the “third generation” of neural network models.
• Neuron models range from the simple leaky integrate-and-fire to the detailed Hodgkin-Huxley equations.
• Training is harder than for standard networks; surrogate gradients, conversion and STDP are the main approaches.
• SNNs run efficiently on neuromorphic computing chips and are simplified models of what living brain organoids do.

How does a spiking neural network work? #

An SNN simulates neurons whose state evolves over time. The most common model is the leaky integrate-and-fire (LIF) neuron.

1. Integrate. Incoming spikes, weighted by synaptic strengths, push the neuron’s membrane potential up (excitatory) or down (inhibitory).
2. Leak. Without input, the potential decays back toward rest.
3. Fire. When the potential crosses a threshold, the neuron emits a spike, a simplified action potential.
4. Reset. The potential resets, often with a short refractory period before the neuron can fire again.
5. Propagate. The spike travels to downstream neurons, sometimes after a delay.

Because neurons only communicate when they spike, activity is often sparse. On event-driven hardware this can save a great deal of energy, since silent neurons do almost no work (Roy, Jaiswal & Panda, 2019).

Which neuron models do SNNs use? #

Izhikevich’s 2003 model in IEEE Transactions on Neural Networks is popular because it reproduces many firing behaviors seen in cortical neurons, such as bursting and adaptation, with only two equations.

Why call SNNs the “third generation”? #

Maass (1997, Neural Networks) offered a useful way to place SNNs in history. First-generation networks used simple threshold units that output either on or off. Second-generation networks, which include most of today’s deep learning, use continuous activation functions and can be trained with gradients. Third-generation networks use spiking neurons, in which the timing of individual spikes can carry information. Maass argued on theoretical grounds that networks of spiking neurons are at least as computationally powerful as earlier generations for some tasks, while matching the way real neurons communicate more closely.

How is information encoded in spikes? #

Turning data into spikes and back is a core design choice, closely related to neural encoding and decoding in neuroscience.

• Rate coding. A stronger input produces a higher firing rate. Simple and robust, but slow, since it needs time to estimate rates.
• Temporal coding. Information is in exact spike times, such as time to first spike. Fast and efficient but sensitive to noise.
• Population coding. Patterns of activity across many neurons represent a value together.
• Event-based sensors. Some cameras and audio sensors output spike-like events directly, which suits SNNs well.

How are spiking neural networks trained? #

Spikes are all-or-nothing, so their derivative is zero almost everywhere. Standard backpropagation does not work directly. Researchers use several workarounds:

1. Local plasticity rules. Spike-timing-dependent plasticity and other forms of Hebbian learning update synapses based on local spike timing, without a global error signal.
2. ANN-to-SNN conversion. Train a conventional network, then map its activations to firing rates in an equivalent spiking network.
3. Surrogate gradients. Replace the spike’s derivative with a smooth stand-in during the backward pass, enabling backpropagation through time. Neftci, Mostafa and Zenke (2019, IEEE Signal Processing Magazine) reviewed this approach.
4. Reservoir methods. Keep a recurrent spiking network fixed and train only a readout, as in the liquid state machine of Maass, Natschläger and Markram (2002). See reservoir computing.

Spiking neural networks vs conventional neural networks vs living networks #

Why do spiking neural networks matter for biocomputing? #

A living culture on a microelectrode array is, in effect, a biological spiking neural network. SNNs matter for organoid intelligence in several ways:

• Interpretation. SNN models help researchers form hypotheses about what recorded spike patterns, including network bursts, mean.
• Experiment design. Simulations can test stimulation and feedback protocols before running them on scarce living tissue.
• Baselines. A fair test of a living computer compares it with an SNN given the same task, inputs and energy budget.
• Shared tooling. Spike data formats, encoders and decoders can be reused across simulated and biological systems.

The energy question is especially important. Our article on biocomputing vs silicon energy efficiency compares the brain’s efficiency with digital hardware, and SNNs on neuromorphic chips are the closest silicon point of comparison. For how living networks adapt through feedback, see how neurons learn in a dish.

What are the limitations of spiking neural networks? #

• Training is still harder than for conventional networks, and accuracy on many benchmarks often trails them.
• Software tools are less mature than mainstream deep learning frameworks.
• Efficiency gains depend on hardware. Simulating SNNs on GPUs can erase the energy advantage.
• Biological realism is limited. Even detailed SNNs omit much of real neuron and synapse biology.

Making spike data from living tissue open and easy to compare with simulations is part of the problem we are working on at Organoid Programming.

Conclusion #

A spiking neural network computes with discrete, timed spikes, borrowing the basic signaling scheme of real neurons. SNNs bridge machine learning, neuromorphic hardware and neuroscience. For biocomputing, the spiking neural network is both a model of the living tissue we work with and the benchmark it has to beat.

Frequently asked questions

What is a spiking neural network in simple terms?

A spiking neural network is a type of artificial neural network that sends information as short pulses, called spikes, instead of continuous numbers. Each simulated neuron collects input until it reaches a threshold, then fires a spike and resets. Because information is carried in when spikes happen, SNNs naturally handle time-varying data and can run efficiently on event-driven neuromorphic chips.

How is a spiking neural network different from a regular neural network?

Regular artificial neural networks pass real-valued activations between layers in a single step and are trained with backpropagation. Spiking neural networks have neurons with internal state that evolves over time and emit binary spike events. This makes them more biologically realistic and potentially more energy efficient on suitable hardware, but harder to train because spikes are not differentiable.

How are spiking neural networks trained?

Common approaches include local plasticity rules such as spike-timing-dependent plasticity, converting a trained conventional network into a spiking one, and surrogate gradient methods. Surrogate gradients replace the non-differentiable spike with a smooth approximation during the backward pass, letting researchers use backpropagation through time. Neftci, Mostafa and Zenke reviewed this approach in IEEE Signal Processing Magazine in 2019.

Why are spiking neural networks called the third generation?

Wolfgang Maass described spiking neuron networks as the third generation of neural network models in a 1997 paper in Neural Networks. In his framing, the first generation used threshold units, the second used continuous activation functions such as sigmoids, and the third used spiking neurons, which can use the timing of individual spikes to carry information.

How do spiking neural networks relate to organoid intelligence?

Brain organoids are biological spiking neural networks. The models used to build SNNs, such as integrate-and-fire neurons and STDP, are simplified descriptions of what real neurons do. SNN tools help researchers simulate, interpret and design experiments on living cultures, and SNNs offer a useful baseline for judging whether a living network computes better or more efficiently than a simulated one.

Sources

1. Maass (1997). Networks of spiking neurons: the third generation of neural network models. Neural Networks.
2. Hodgkin & Huxley (1952). A quantitative description of membrane current and its application to conduction and excitation in nerve. The Journal of Physiology.
3. Izhikevich (2003). Simple model of spiking neurons. IEEE Transactions on Neural Networks.
4. Neftci, Mostafa & Zenke (2019). Surrogate gradient learning in spiking neural networks. IEEE Signal Processing Magazine.
5. Roy, Jaiswal & Panda (2019). Towards spike-based machine intelligence with neuromorphic computing. Nature.
6. Maass, Natschläger & Markram (2002). Real-time computing without stable states: a new framework for neural computation based on perturbations. Neural Computation.