A neuron model is a mathematical or computational description of how a nerve cell receives, processes, and transmits electrical signals. The most celebrated version, the Hodgkin-Huxley model developed in 1952, uses a set of differential equations to reproduce the voltage spikes that carry information through your brain and body. Since then, researchers have built dozens of model variants at different levels of detail, and these models now underpin everything from the artificial neural networks running modern AI to brain implants that let paralyzed patients control robotic arms.
What a Real Neuron Actually Does
Before you can appreciate the model, it helps to know the basics of the real thing. A neuron at rest maintains a small voltage difference across its outer membrane, roughly negative 70 millivolts on the inside relative to the outside. That resting voltage exists because the membrane is selectively permeable to certain ions, particularly potassium and sodium, and ion pumps actively maintain concentration gradients across it. The quantitative relationship between these ion gradients and the membrane’s permeability to each ion determines the resting potential.
When a neuron receives enough excitatory input from its neighbors, voltage-gated sodium channels in its membrane snap open. Sodium ions rush inward, the interior voltage shoots upward, and an action potential fires. This is the fundamental electrical spike that carries information. Hodgkin and Huxley first described this sodium-dependent activation, along with the rapid inactivation that follows, in their landmark series of papers in 1952.
Most neurons also have elaborate branching structures called dendrites, which collect incoming signals. These dendrites are not simple wires. The way electrical signals travel through them depends on properties like how membrane conductance varies along the branch. Analytical studies have shown that even keeping the total number of ion channels fixed, changing how those channels are distributed along a dendrite alters how efficiently voltage signals reach the cell body, how quickly transients decay, and how the system responds to synaptic inputs versus injected current.
Communication between neurons is usually described as happening at synapses, either through chemical neurotransmitters or through direct electrical connections called gap junctions. But neurons can also influence each other through a subtler mechanism called ephaptic coupling, where the extracellular voltage changes generated by one neuron’s action potential are large enough to open sodium channels in a neighboring neuron’s axon. Research on cerebellar Purkinje cells has demonstrated that this mechanism promotes synchronized firing between nearby cells, entirely independent of chemical or electrical synapses.
Neurons do not operate in isolation, either. Astrocytes, the star-shaped glial cells once thought to be mere structural support, actively influence neural network dynamics. They integrate neural activity across local circuits and can drive bursts of action potentials in certain populations of inhibitory neurons, shaping the timing and spatial patterns of brain activity.
The Hodgkin-Huxley Model
The foundational neuron model emerged from experiments on the giant axon of the squid, which is large enough to thread electrodes inside. Using a technique called voltage clamping, Alan Hodgkin and Andrew Huxley measured how sodium and potassium currents change with voltage and time, then translated those measurements into equations. Their model treats the neuron’s membrane as an electrical circuit: a capacitor in parallel with conductances for sodium, potassium, and a leak current, each with its own voltage-dependent gating variables. Feed a stimulus current into this circuit and the equations produce a voltage spike that looks remarkably like a real action potential.
The model’s achievement was not just qualitative resemblance. It quantitatively predicted the speed of action potential propagation, the refractory period during which a neuron cannot fire again, and the threshold behavior where sub-threshold stimuli produce nothing while super-threshold stimuli produce a full spike. Hodgkin and Huxley shared the 1963 Nobel Prize for this work, and the model has endured for decades as the gold standard for biophysically detailed simulation.
That said, the Hodgkin-Huxley model is computationally expensive. Each neuron requires solving several coupled differential equations at every time step. Researchers have explored ways to speed things up without losing too much accuracy. One approach uses a simplified noise algorithm based on a Langevin description, which can reduce simulation time by about two orders of magnitude compared to a full master-equation treatment. Another recent line of work extends the model with fractal-fractional operators that capture memory effects and multi-scale dynamics, allowing researchers to reproduce complex firing patterns like intrinsic bursting that the classical equations handle less naturally.
Machine learning has also entered the picture. One study trained a simple artificial neural network to predict specific features of Hodgkin-Huxley spikes, including maximum voltage, minimum voltage, and the interval of the voltage drop, across nine different model variants whose firing patterns cover most behaviors observed in the brain. The approach works as a fast approximation when running the full equations would be too slow.
Simpler Models That Trade Detail for Speed
Not every application needs the biophysical richness of Hodgkin-Huxley. When you want to simulate thousands or millions of neurons interacting in a network, you need something leaner. The most popular alternative is the leaky integrate-and-fire model. It treats the neuron as a capacitor that charges up as it receives input and “leaks” charge over time. When the voltage crosses a threshold, the model registers a spike and resets. There is no attempt to reproduce the detailed shape of the action potential; the model simply marks that a spike occurred and moves on.
What this family of models lacks in biological detail, it gains in analytical tractability. Researchers have established a precise mathematical mapping between the parameters of a linear leaky-integrate-and-fire spiking neural network and the parameters of a conventional deep neural network using a common activation function. This connection is more than a curiosity. It means insights from the deep learning world can be translated into the spiking network world and vice versa, bridging two communities that historically developed their tools independently.
Between these extremes sits a spectrum of models with varying complexity. The Izhikevich model, for instance, uses just two equations but can reproduce over twenty distinct firing patterns seen in real cortical neurons by adjusting four parameters. The FitzHugh-Nagumo model reduces the Hodgkin-Huxley framework to two variables, capturing the essential excitable dynamics while being simple enough to analyze on paper. Which model you choose depends on the question you are asking: if the shape of the spike matters, you lean toward Hodgkin-Huxley; if you care about network-level phenomena like oscillations or synchronization, a simpler model that lets you scale up is often the better tool.
How Biological Neuron Models Inspired Artificial Neural Networks
The artificial neurons in modern AI are distant descendants of biological neuron models, though the family resemblance has faded considerably. The original artificial neuron, the Perceptron, was introduced in the late 1950s as a simplified version of how a nerve cell sums its inputs and fires or stays silent. It multiplied each input by a weight, added the results, and passed the sum through a step function. This was crude but showed that a network of such units could learn to classify patterns.
The Perceptron had serious limitations, and neural network research stalled for years until a nonlinear network algorithm overcame many of them, sparking the resurgence of the field. That algorithm, backpropagation applied to multi-layer networks, required replacing the step function with smooth, differentiable activation functions. Early choices like the logistic sigmoid loosely echoed the S-shaped input-output curves of real neurons. Later, the ReLU function (which simply outputs zero for negative inputs and the input itself for positive ones) became dominant for purely practical reasons: it trains faster and avoids certain numerical problems. A systematic overview of activation functions notes that new variants have “mushroomed” alongside the growth of deep learning, creating genuine confusion about which to use and when.
The gap between artificial and biological neurons has become a topic of scientific debate in its own right. Conventional artificial neural networks operate on firing rates: each unit outputs a single number representing how active it is. Real neurons communicate through discrete spikes whose precise timing may carry information. An epistemological analysis of this question argues that the rate-based view, while practical, has “virtually no empirical or theoretical support” as a description of what actually drives neural activity, and that the causal role of spike timing has been obscured by framing the question from an external observer’s perspective.
Neuromorphic Hardware and Spiking Networks
If biological neurons communicate through spikes rather than smooth rates, can we build hardware that works the same way? Neuromorphic engineering tries to do exactly that. Inspired by the brain’s massive parallelism and sparse coding, neuromorphic chips implement spiking neural networks directly in silicon. Several large-scale projects have demonstrated systems with millions of neurons and billions of synapses, processing sensory information with dramatically lower power consumption than conventional processors.
The energy advantage comes from event-driven processing. In a conventional chip, every unit computes at every time step whether or not anything interesting is happening. In a spiking network, computation happens only when a spike arrives, and most neurons are silent most of the time. A recent comparative review confirms that spiking neural networks use bio-inspired, event-driven architectures that can be significantly more energy-efficient, though their training tools remain less mature than those available for conventional deep networks.
The energy picture is not as simple as “spikes always win,” though. An analytical study of spiking network efficiency found that for an SNN to beat an optimized conventional network, its time window should generally stay below four timesteps with a spike rate below about 7%. Beyond those operating points, a well-optimized quantized conventional network is often the more efficient choice. The takeaway: energy savings depend heavily on co-designing the algorithm and the hardware together, not on spiking dynamics alone.
One promising hardware technology for neuromorphic systems is the memristor, a circuit element whose resistance changes based on the history of current that has flowed through it, mimicking the way synapses strengthen or weaken with use. Researchers have used memristive devices to implement both the nonlinear ion channels of a neuronal oscillator and the synaptic learning rules that adjust connection strengths. In one implementation, the charge accumulation effect in a memristive device produced a biologically plausible synapse that showed potentiation, with increasing response amplitude triggered by a sequence of spikes. Another hardware design combines CMOS circuits for the neuron body with memristors for spike-timing-dependent synapses, creating compact hybrid circuits.
Digital implementations are advancing too. A recent FPGA implementation of a complete digital spiking neuron introduced a power-of-two-based approximation that sidesteps the computational cost of the nonlinear functions (quadratic terms and multiplications) that make traditional digital neuron models expensive.
Medical Applications of Neuron Models
Some of the most consequential applications of neuron models are in medicine, particularly in understanding and treating epilepsy. Seizures involve abnormal, synchronized firing across large populations of neurons, and computational models let researchers dissect the mechanisms driving that synchronization in ways that experiments on living tissue cannot easily achieve. A review of computational epilepsy modeling notes that substantial progress has been made at levels ranging from the molecular to the network scale. One study developed a spiking network model and systematically investigated the biophysical conditions under which it reproduces the dynamic behaviors known from established abstract seizure models, bridging the gap between single-neuron physiology and the emergent patterns clinicians see on an EEG. Large-scale models have also been used to understand how the network and cellular changes induced by epilepsy contribute to cognitive deficits, such as impaired spatial processing linked to disrupted theta oscillations.
Brain-machine interfaces represent another frontier. Here the challenge is decoding the electrical activity of real neurons in real time and translating it into commands for a prosthetic limb or computer cursor. A study demonstrated that a Kalman-filter-based decoder, implemented as a spiking neural network of about 2,000 neurons using a general framework for mapping control algorithms onto SNNs, performed comparably to a standard floating-point Kalman filter in closed-loop experiments with two rhesus monkeys. The significance is practical: spiking network decoders could eventually run on ultra-low-power neuromorphic chips implanted directly in the brain, eliminating the need for power-hungry conventional processors connected by wires through the skull.
Bidirectional neural interfaces push this idea further. Rather than just reading from the brain, these systems also write back into it. A proof-of-concept system demonstrated that a bidirectional interface connecting two neural networks in a closed loop could establish and control network properties like synchrony significantly more effectively than unidirectional alternatives or no interface at all. This kind of technology points toward future prosthetics that not only obey your commands but also feed sensory information back into your nervous system.
Predictive Coding and How the Brain Sees
Neuron models are also reshaping theories of perception. Predictive coding is the idea that your brain does not passively absorb sensory information but actively generates predictions about what it expects to sense, then mainly processes the prediction errors, the differences between what it expected and what actually arrived. Implementing this idea in spiking neural networks has produced models that perform perceptual inference and learning in an unsupervised manner, using biologically plausible mechanisms.
A recent model called Predictive Coding Light takes an interesting twist on the standard framework. Instead of transmitting prediction errors up to higher processing stages, it suppresses the most predictable spikes and transmits a compressed representation of the input. Using only spike-timing-based learning rules, it reproduces a range of findings on information processing in visual cortex and achieves strong performance on image classification tasks. The approach suggests that the brain’s sensory processing may be even more efficient than standard predictive coding theories imply, sending less information up the hierarchy rather than more.
Learning Rules in Spiking Networks
A model neuron that cannot learn is of limited use. In biological brains, synaptic connections strengthen or weaken depending on the relative timing of spikes in the connected neurons, a phenomenon known as spike-timing-dependent plasticity. But timing alone is not enough to explain how the brain learns to perform useful tasks. A third factor is needed: a neuromodulatory signal, like dopamine, that tells the network whether a recent action led to reward or punishment.
In reward-modulated spike-timing-dependent plasticity, the connection between two neurons changes only when a neuromodulatory signal indicates that something noteworthy happened. The synapse maintains an eligibility trace, essentially a short-term memory of recent spike coincidences, and the actual weight change occurs when that trace is multiplied by the reward signal. This three-factor rule solves a fundamental problem: how does a synapse deep inside the brain know whether the organism’s behavior was good or bad? The reward signal broadcasts that information globally, and only synapses that recently participated in producing the behavior get updated. This framework has become central to building spiking networks that can learn from interaction with an environment rather than from labeled training data.
When Models Meet Living Tissue
Perhaps the most provocative development in neuron modeling is the emergence of hybrid systems that connect computational models directly to living biological neurons. The DishBrain project, which cultured human neurons on a multi-electrode array and taught them to play a simple video game, attracted worldwide attention and immediate ethical scrutiny. The creation of such hybrid, silico-biological intelligence raises concerns that go beyond standard AI ethics.
Following the neuroscientific framework embraced by the DishBrain creators themselves, ethicists have highlighted how the system’s design, grounded in the Free Energy Principle, may risk bringing about artificial suffering, and have argued for a cautious approach to such synthetic biological intelligence. A broader analysis categorizes the ethical concerns into two groups: those related to the moral standing of the hybrid system itself (since it contains human neural components that might exhibit properties indicative of moral standing) and those related to the cell donors, including questions of consent, responsibility for the system’s behavior, and benefit sharing. These are not distant hypotheticals. As neuron models become sophisticated enough to interface with real tissue in closed loops, and as the tissue itself grows complex enough to exhibit learning, the line between simulation and experience becomes genuinely unclear.
Neurons Without Nervous Systems
The neuron model story has an unexpected evolutionary footnote. Placozoans, among the simplest multicellular animals on Earth, lack anything resembling a nervous system. They are tiny, flat creatures that glide along surfaces without muscles, nerves, or synapses. Yet researchers have shown that four species of placozoans produce rapid, sodium-dependent action potentials, the same fundamental electrical signal that the Hodgkin-Huxley model was built to describe. Molecular analysis revealed five to seven different types of voltage-gated sodium channels in these animals, a surprising diversity that reflects both gene duplication events and parallel evolution. The finding suggests that the electrical signaling machinery modeled by Hodgkin and Huxley is far older and more widespread than neurons themselves, serving as a tool for behavioral coordination long before anything like a brain existed.