What Is a Spiking Neural Network?

Your Brain Doesn't Do Math the Way Computers Do. And That Might Be the Whole Problem.

Here's a question that should bother you more than it probably does: the most powerful AI systems on Earth, the ones writing poetry, generating images, and beating grandmasters at chess, are called "neural networks." They're named after your brain. But they work almost nothing like your brain.

A real neuron in your head communicates by firing brief electrical pulses. It sits quietly, accumulating charge, and then, when the pressure crosses a threshold, it fires a spike. A single, sharp burst of electricity lasting about a millisecond. Then it goes quiet again. The information isn't just in whether it fires. It's in when it fires. The precise timing of that spike relative to everything else happening in the neural neighborhood. Your brain runs on timing.

The "neural networks" powering ChatGPT, image generators, and self-driving cars? They don't do any of that. They pass continuous numbers through mathematical functions. No spikes. No timing. No silence between events. Every artificial neuron in every layer crunches a number on every single computation cycle, regardless of whether it has anything useful to contribute.

It's a bit like saying you built a "bird-inspired flying machine" that has no wings, doesn't flap, and runs on jet fuel. Sure, it flies. But the bird would not recognize it as a relative.

Spiking neural networks are the other approach. The one that actually took the "neural" part seriously. And after decades of living in the shadow of conventional AI, they're becoming impossible to ignore.

The 30-Second Version (Then We'll Go Deep)

A spiking neural network (SNN) is an artificial neural network where the artificial neurons communicate using discrete electrical pulses, called spikes, that occur at specific moments in time. Instead of passing continuous numbers between layers the way a traditional neural network does, an SNN's neurons accumulate input, fire a spike when a threshold is crossed, then reset. The information lives in the timing and pattern of those spikes, not just their presence or absence.

This is a fundamentally different model of computation. And it happens to be the model your brain has been using for roughly 500 million years.

First, Why Traditional Neural Networks Are "Neural" in Name Only

To understand why spiking neural networks matter, you need to understand what they're rebelling against.

A conventional artificial neural network, the kind that powers nearly all modern AI, works like this. You have layers of artificial neurons. Each neuron receives numerical inputs, multiplies each one by a weight, sums them up, pushes the total through a mathematical function (called an activation function), and outputs a single number. That number flows to the next layer. Repeat until you get an answer.

This process happens in perfectly synchronized lockstep. Every neuron in layer one computes its output. Then every neuron in layer two computes its output using layer one's results. Then layer three. It's orderly, parallel, and completely unlike anything happening in your skull.

Here's what's missing:

Time. In a conventional neural network, there is no inherent concept of time. Inputs are static snapshots. If you want to process something that changes over time, like speech, video, or brainwaves, you have to chop it into frames, convert each frame into a static vector of numbers, and process them one chunk at a time. The temporal structure of the original signal gets flattened.

Silence. In a conventional neural network, every neuron fires on every pass. There's no concept of a neuron being quiet, waiting, or choosing not to contribute. This means the network is always doing maximum work, even when most of that work is irrelevant.

Spike timing. Your biological neurons encode information in the precise millisecond timing of their spikes. Two neurons firing 5 milliseconds apart carry different information than the same two neurons firing 20 milliseconds apart. Conventional neural networks have no mechanism for this. Their "neurons" output a single floating-point number, and the timing of computation is dictated by the clock cycle of the processor, not by the information itself.

Spiking neural networks bring all three of these things back.

How Does a Spiking Neuron Actually Work?

A spiking artificial neuron is modeled, at least loosely, on the dynamics of a real biological neuron. The most common model is called the leaky integrate-and-fire (LIF) neuron, and it works like this.

Imagine a bucket with a small hole in the bottom. Water flows in from various sources (these are the input signals from other neurons). The water level in the bucket represents the neuron's membrane potential, its accumulated electrical charge. The hole in the bottom means the bucket is constantly leaking. Charge slowly drains away over time, just like a real neuron's membrane potential decays between inputs.

If enough water flows in fast enough to overcome the leak, the water level hits the rim. The bucket overflows. That overflow is the spike: a sudden, discrete event. After the spike, the bucket gets emptied (the membrane potential resets to a resting value), and there's a brief refractory period where no amount of input can make it fire again. Then the process starts over.

This is astonishingly close to what an actual neuron does. Real neurons accumulate charge from incoming postsynaptic potentials. The charge leaks through ion channels. If the membrane potential reaches roughly negative 55 millivolts (the threshold), the neuron fires an action potential. Then it resets.

Two Languages of the Brain: Rate Codes vs. Temporal Codes

This brings us to one of the most fascinating debates in neuroscience, one that spiking neural networks force us to confront directly.

For decades, the dominant theory of neural coding was rate coding. The idea is simple: what matters is how many spikes a neuron fires per second. A neuron firing at 80 spikes per second is saying "strong signal." A neuron firing at 20 spikes per second is saying "weak signal." The precise timing of individual spikes? Just noise. Doesn't matter.

Rate coding is clean, intuitive, and relatively easy to work with mathematically. It's also basically what conventional neural networks implement. That single floating-point number output by each artificial neuron? It's analogous to a firing rate.

But here's where it gets interesting. Starting in the 1990s, mounting evidence suggested that the brain uses far more than just rate codes. Researchers found that neurons in the visual cortex could distinguish between stimuli based on the timing of their very first spike, long before enough spikes had accumulated to estimate a rate. In the auditory system, neurons encode the location of sounds using timing differences of just microseconds. In the hippocampus, the precise phase at which a neuron fires relative to the background theta rhythm encodes an animal's position in space.

This is temporal coding. And it implies something profound: the brain is doing computation with time itself. The timing of each spike isn't noise to be averaged away. It's signal. It's data. It's part of the answer.

Spiking neural networks are the first artificial systems that can naturally implement temporal coding. And this opens doors that rate-based systems fundamentally cannot walk through.

The "I Had No Idea" Moment: Your Brain Is Absurdly Energy-Efficient

Here's the fact that rearranges everything.

Your brain consumes about 20 watts of power. That's roughly the same as a dim light bulb. With those 20 watts, it runs 86 billion neurons, manages roughly 100 trillion synaptic connections, processes visual scenes in real time, understands language, generates emotion, maintains consciousness, and keeps you breathing while doing all of it.

GPT-4, by contrast, is estimated to require a cluster of around 25,000 GPUs for training, with each GPU drawing 300 to 700 watts. During inference (actually generating text), a single query reportedly consumes around 10 times more energy than a Google search. The total power budget for running a large language model at scale is measured in megawatts.

Your brain does something comparably sophisticated, arguably more sophisticated, on 20 watts. That's not a small difference. That's a difference of roughly five to six orders of magnitude.

How? Spikes.

The secret is event-driven computation. In your brain, the vast majority of neurons are silent at any given moment. They're not using energy. They're not computing anything. They're just sitting at their resting potential, waiting. Only when they receive enough input to cross threshold do they fire a spike and briefly consume energy. This sparse, on-demand activation means the brain only spends energy where and when it's needed.

Traditional neural networks do the opposite. Every neuron activates on every forward pass. If you have a network with 175 billion parameters (like GPT-3), then 175 billion multiply-accumulate operations happen for every single token you generate. Most of that computation is probably irrelevant for any given token, but the architecture doesn't know that. It computes everything every time.

Spiking neural networks inherit the biological advantage. Neurons that have nothing to contribute remain silent and consume no energy. Only the relevant neurons fire. Researchers have demonstrated SNNs performing image classification tasks at 100 to 1,000 times lower energy than equivalent conventional networks, and that's on hardware that wasn't even designed for them.

Brainwave data, captured at 256Hz across 8 channels, processed on-device. The Crown's open SDKs let developers build brain-responsive applications.

Explore the Crown

Neuromorphic Hardware: Chips That Think in Spikes

Here's the catch with spiking neural networks. If you try to simulate them on a conventional computer, you lose most of the efficiency advantage. A regular CPU or GPU processes everything in synchronized clock cycles. Simulating the asynchronous, event-driven behavior of spiking neurons on synchronous hardware is like simulating ocean waves using a spreadsheet. You can do it, but you're fighting the architecture.

This is why neuromorphic hardware exists. These are computer chips designed from the ground up to run spiking neural networks natively.

The most prominent examples:

Look at those power numbers. Intel's Loihi 2 runs a million artificial spiking neurons on less than a watt. IBM's TrueNorth chip, released back in 2014, ran a million neurons on 70 milliwatts. That's roughly the power consumption of a hearing aid.

Compare that to training a modern large language model, which can consume the energy equivalent of hundreds of homes over several months, and you start to see why neuromorphic computing isn't just an academic curiosity. It's potentially the only path to putting truly intelligent systems into devices that run on batteries.

How SNNs Learn: STDP and the Timing of Change

Training a spiking neural network is one of the hardest open problems in computational neuroscience. And the reason is both elegant and frustrating.

In conventional neural networks, training relies on backpropagation: you compute how wrong the output is, then send error signals backward through the network, adjusting each weight by a precise amount. This requires computing gradients, the mathematical derivatives of the error with respect to each weight. And it works beautifully because the activation functions in conventional neural networks are smooth, continuous, and differentiable.

Spikes are none of those things.

A spike is a binary event. It either happens or it doesn't. There's no "half spike." The mathematical function that describes "fire a spike when the membrane potential crosses threshold" is a step function, and step functions have a gradient of zero everywhere except at the exact threshold, where the gradient is undefined. You can't backpropagate through undefined.

So how do spiking networks learn?

The biological answer is spike-timing-dependent plasticity (STDP). Discovered experimentally in the late 1990s, STDP is a learning rule based on a beautifully simple principle: if neuron A fires just before neuron B, the connection from A to B gets stronger. If neuron A fires just after neuron B, the connection gets weaker.

Think about what this means. The brain is essentially saying: "If A's spike helped cause B's spike (because A fired first), then that connection is useful, so strengthen it. If A fired after B had already fired (so A's spike couldn't have caused B's spike), weaken the connection."

The timing window is narrow, roughly 10 to 20 milliseconds. This creates a Hebbian learning rule with a temporal twist: neurons that fire together wire together, but only if they fire in the right order.

STDP is unsupervised. It doesn't need labeled training data or an external error signal. It just watches the patterns of spikes flowing through the network and adjusts connection strengths based on timing correlations. It's the learning rule that built your brain.

For engineering applications, researchers have also developed surrogate gradient methods. The idea is clever: since you can't compute the true gradient of a spike, you approximate it with a smooth function that has a similar shape. During the forward pass, spikes are binary. During the backward pass, you pretend the spiking function was a smooth sigmoid or similar curve, and compute gradients through that. It's mathematically impure but practically effective, and it's allowed SNNs to achieve competitive performance on benchmarks like MNIST and CIFAR-10.

Why Brain-Computer Interfaces Need Spiking Neural Networks

Here's where all of this converges on something deeply practical.

When you put an EEG headset on your head, the signals you're recording are the aggregate electrical activity of millions of neurons that communicate through spikes. The oscillatory patterns that EEG captures (alpha brainwaves, beta brainwaves, theta, gamma) emerge from the synchronized spiking of large neural populations. The raw data is inherently temporal. It unfolds over time. The information isn't just in the amplitudes. It's in the precise timing of peaks and troughs, in the phase relationships between oscillations at different frequencies, in the millisecond-level dynamics that change from moment to moment.

Now consider what happens when you process EEG data with a conventional neural network. You chop the continuous signal into windows. You extract features: power spectral density, mean amplitude, standard deviation. You flatten the temporal structure into a feature vector, a static snapshot. Then you feed it into a network that has no native concept of time.

You're throwing away exactly the information that the brain encoded most carefully.

Spiking neural networks offer a fundamentally different approach. Because they operate in continuous time, they can process EEG data as it arrives, spike by spike, sample by sample, without chopping it into windows. The temporal relationships in the data are preserved natively. Phase relationships between oscillations, millisecond-level shifts in timing, subtle temporal patterns that distinguish focused attention from mind-wandering: an SNN can represent all of these in its own spiking dynamics.

The energy advantage is especially important for wearable BCIs. A device you wear on your head all day can't afford to run power-hungry GPU inference. But a neuromorphic chip running an SNN? That could run for days on a watch battery.

And the latency advantage matters too. In a brain-computer interface, the difference between 200 milliseconds of processing delay and 10 milliseconds can be the difference between an interface that feels like an extension of your body and one that feels laggy and disconnected.

The Frontier: What Comes Next

Spiking neural networks are not yet at parity with conventional deep learning on most benchmarks. The training algorithms are harder. The software toolchains are less mature. The hardware, while advancing rapidly, is still generations behind the GPU ecosystem regarding availability and developer support.

But the trajectory is clear. Intel's Loihi 2 is on its second generation. BrainChip's Akida is being embedded in edge devices for always-on sensory processing. The open-source frameworks for SNN development (Norse, snnTorch, Lava) are gaining contributors and momentum.

And the applications that benefit most from SNNs, the ones where temporal precision, energy efficiency, and biological compatibility matter most, are exactly the applications that matter for the future of brain-computer interfaces.

Consider what becomes possible when the neural decoder in a BCI is itself a spiking neural network. You'd have a system where biological spikes (from the user's brain) are translated into artificial spikes (in the decoder), processed through dynamics that mirror the brain's own computational principles, and converted into commands or feedback with single-digit millisecond latency, all on a chip consuming less power than a Bluetooth radio.

That's not science fiction. Every piece of that pipeline exists today in some form. The work now is integration, optimization, and scale.

The Deeper Thought: Maybe the Brain Was the Blueprint All Along

There's a certain irony in the history of artificial intelligence. We started by trying to build systems inspired by the brain. The first artificial neuron, the McCulloch-Pitts model from 1943, was explicitly based on biological neurons. It even fired in a binary, all-or-nothing fashion, just like a real spike.

Then we drifted. Backpropagation, continuous activations, synchronized layer-by-layer processing. These innovations made neural networks trainable and powerful, but they moved further and further from biology with each generation. By the time we reached modern large language models with billions of parameters running on football-field-sized data centers, the connection to actual neurons was purely etymological.

Spiking neural networks represent a return. Not a nostalgic one, not a rejection of everything deep learning has accomplished, but a recognition that maybe the brain knew something we forgot. Maybe the reason a 20-watt organ outperforms a megawatt data center on so many cognitive tasks is that evolution spent 500 million years optimizing an architecture that we've barely begun to understand.

The EEG signals you can measure right now, sitting at your desk, are the surface echoes of that architecture at work. Every alpha wave, every burst of gamma activity, every shift in the spectral profile of your brainwaves reflects the coordinated spiking of millions of neurons processing information in time, using codes we're only starting to decode.

Understanding spiking neural networks isn't just an exercise in computer science. It's a window into the computational language of the brain itself. And if we can learn to speak that language, to build machines that compute the way biology computes, we won't just build better AI. We'll build machines that can finally talk to the brain in its native tongue.

That conversation has already started. It just needs better translators.