Spiking Neural Networks in Advanced Brain Studies

Spiking neural networks are computational models that mimic the brain’s own signaling language: brief electrical pulses, or spikes, fired between neurons. Unlike conventional artificial neural networks, which pass continuous numbers between layers, spiking neural networks (SNNs) encode information in the precise timing and patterns of discrete spikes, much the way real neurons do. That biological fidelity has made them one of the most versatile tools in modern brain research, used to reconstruct cortical circuits, decode brain signals for prosthetic limbs, simulate the effects of neurological disease, and even test theories of consciousness.

Why Spikes Matter More Than Averages

Real neurons do not communicate through smooth, continuously varying signals. They accumulate incoming electrical charge until a threshold is reached, fire a spike, then briefly go silent during a refractory period before the cycle begins again. Each of the mathematical neuron models used in SNNs reproduces this basic loop, though they differ in how they handle the details. Some generate simple, regular spike trains; others exhibit complex patterns such as bursting or delayed responses, depending on their internal dynamics and parameter settings.

This spike-by-spike operation gives SNNs a closer correspondence to brain activity than traditional deep-learning networks, which treat a neuron’s output as a single number summarizing its average firing rate. In the brain, two neurons can fire at the same average rate yet carry very different information depending on when their spikes land relative to each other and to an external stimulus. SNNs capture that temporal dimension, which is why researchers use them when the question being asked is not just “what does the brain compute?” but “how and when does it compute it?”

Neural Coding Schemes and Their Trade-Offs

One of the first decisions when building an SNN is choosing how information gets encoded in spikes. The two most common approaches are rate coding, where information is carried by how many spikes a neuron fires per unit of time, and time-to-first-spike (TTFS) coding, where the key variable is how quickly a neuron fires after receiving input. Rate coding is intuitive and robust to noise, but it is slow and computationally expensive because you need many spikes to estimate a rate reliably.

TTFS coding turns out to be far more efficient. Comparative benchmarks show that TTFS coding can achieve the highest computational performance while requiring roughly four times less processing latency and about three-and-a-half times fewer spike operations than rate coding during training, with even larger savings during inference.¹ That efficiency matters both for brain simulation, where each saved operation means a larger network can be simulated, and for neuromorphic hardware destined for real-world devices like brain-computer interfaces.

Modeling Cortical Circuits at Full Scale

A central ambition in computational neuroscience is to build spiking models of the cortical microcircuit, the small patch of neural tissue that repeats across the brain’s surface like a tile pattern and is thought to perform a canonical computation. A landmark full-scale spiking model of this microcircuit showed that when the connectivity map was drawn from anatomical and physiological data, the simulated network spontaneously produced the kind of irregular, asynchronous firing seen in awake animals, including the characteristically low firing rates of excitatory cells in the upper cortical layers.² Getting that behavior to emerge, rather than having to impose it artificially, was a strong sign that the model had captured something real about how cortical circuits balance excitation and inhibition.

Running these models is computationally demanding. A cortical microcircuit containing roughly 77,000 neurons and 300 million connections can now be simulated close to real time on a single modern GPU. Scaling up to a million neurons with a thousand connections each pushes the simulation to about 70 seconds of compute per second of biological activity.³ That is still far from real time for million-neuron models, but it represents a dramatic improvement over what was possible just a decade ago, and it opens the door to experiments that would be impractical or impossible to conduct in living tissue.

Reconstructing Sensory Systems

Vision and hearing are among the brain’s most heavily studied sensory systems, and SNNs have become a primary tool for reconstructing how they work. In vision research, spiking models have been used to replicate how the retina processes incoming light before signals ever reach the brain. A model optimized for efficient temporal prediction was found to develop receptive fields that look like those in real retinal ganglion cells, and it captured phenomena such as latency coding, response omissions when expected stimuli fail to appear, and sensitivity to the direction of motion. Strikingly, the same model also predicted actual ganglion-cell responses to natural images and movies across multiple animal species more accurately than prior approaches.⁴

Hardware-oriented work has pushed this further. One project built a modular organic neuromorphic circuit that pairs spiking neurons with biohybrid synapses mediated by actual neurotransmitters. The spiking neuron converts light stimuli into coded spike patterns, mimicking an afferent neuron, while the biohybrid synapse recreates the modulatory role of interneurons. Together, these components form a retina-inspired pre-processing pathway on a chip.⁵ Beyond vision, SNNs have also been used to model human-like auditory pathways, integrating audiovisual information in ways that parallel multisensory processing in the brain.⁶

A practical twist on sensory-system research involves event-driven cameras, sometimes called dynamic vision sensors, that output spikes rather than frames. A recognition system that feeds these spike trains into a three-dimensional spiking neural network modeled on the organization of the primary visual cortex has demonstrated accurate identification of moving objects while preserving the temporal resolution that conventional frame-based cameras lose.⁷

Dendrites as Computational Units

Most SNNs treat each neuron as a simple point that adds up its inputs and fires when a threshold is crossed. Real neurons are far more elaborate. Their dendrites, the tree-like branches that receive input from other neurons, perform their own nonlinear computations before signals ever reach the cell body. A growing body of work is showing that ignoring dendrites leaves significant computational power on the table.

One line of research introduced modules that emulate two dendritic functions: spatial gating, which scales and shifts inputs to reduce the information lost when a continuous membrane voltage is converted into a binary spike, and temporal adjustment, which lets the dendrite weight inputs from different time steps differently. Combining these two mechanisms produced a more balanced distribution of spikes across the network and measurably improved performance on benchmark tasks.⁸ Separately, a framework called Dendrify allows researchers to build reduced compartmental neuron models that include phenomenological voltage-gated mechanisms for dendrites, replicating various dendritic features and their effects on somatic output while keeping the computational overhead manageable enough for large-scale network simulations.⁹

Theoretical work has demonstrated that even passive, conductance-based dendritic interactions can be systematically exploited to compute a wide variety of multivariate functions, including operations like the Euclidean norm and nonnegative multiplication, that are difficult for simple point neurons.¹⁰ These findings suggest that the brain’s computational power may reside partly in the dendrites themselves, not just in the connections between neurons, and that future SNNs incorporating dendritic structure could be both more biologically accurate and more capable.

Brain-Computer Interfaces

SNNs are a natural fit for brain-computer interfaces (BCIs), which decode neural activity to control a prosthetic limb or a cursor on a screen. The brain itself speaks in spikes, so an SNN decoder can work in the same language as its input. An early demonstration showed that a 1,600-neuron SNN could implement a Kalman filter, one of the most widely used decoding algorithms in high-performance neural prostheses, and match the performance of a conventional implementation while running in real time on a standard desktop computer.¹¹

More recent work has tackled a harder problem: keeping a BCI accurate when the neural signals it reads inevitably drift over time due to electrode movement, tissue changes, or shifts in the patient’s neural state. A continuous-learning SNN achieved the same peak decoding accuracy as existing offline-trained SNNs on real neural data while cutting total memory usage by 92 percent. It also matched the best available methods in a closed-loop setting, adapted successfully to multiple types of input disruption, and could be trained entirely online, with no prior offline training session required.¹² That last point matters for clinical use, because an interface that must be retrained from scratch every time something shifts is impractical for daily life.

Disease Modeling and Deep Brain Stimulation

When a brain circuit misfires, the consequences can range from tremor to seizure to cognitive decline. SNN models of diseased circuits offer a way to study these breakdowns at a resolution that brain imaging alone cannot provide. A large-scale spiking model of the basal ganglia, the deep-brain circuit involved in movement control and reward learning, was tuned to reproduce the firing patterns observed in both brain-slice recordings and in vivo behavioral studies.¹³ Models like this allow researchers to manipulate specific cell types or connection strengths and observe how the disruption propagates, something that is extremely difficult to do cleanly in a living brain.

SNN-based disease models are also beginning to inform treatment. Deep brain stimulation (DBS), where an implanted electrode delivers electrical pulses to specific brain regions, is already used for Parkinson’s disease and treatment-resistant depression. Most current DBS systems deliver stimulation at fixed parameters regardless of what the brain is doing at any moment. A closed-loop approach, where stimulation adapts in real time to the brain’s current state, could be more effective and more energy-efficient. Researchers have proposed using a spiking reinforcement-learning agent, specifically a deep spiking Q-network, to optimize DBS parameters on the fly within a simulated basal ganglia circuit, validating an end-to-end spike-based control model.¹⁴

At a broader scale, there is active work on using neuromorphic computing to simulate the neural systems affected by conditions like Alzheimer’s, Parkinson’s, and epilepsy. The goal is to use the detailed temporal dynamics of spiking models to optimize drug discovery pipelines, for example by predicting whether a candidate compound will restore normal firing patterns in a diseased circuit before running expensive animal trials or human studies.¹⁵

Mapping the Whole Brain with fMRI-Constrained Networks

Most of the models discussed so far focus on a single brain region or circuit. But many of neuroscience’s biggest questions involve the interactions between regions, the kind of large-scale functional architecture that brain imaging techniques like fMRI reveal. A newer approach constrains an SNN’s wiring by the functional connectivity patterns measured in human fMRI data, building a network whose long-range connections mirror the ones observed in a living brain. The resulting fMRI-constrained SNN can then be tested for properties such as resilience to random damage, offering a way to study how the brain’s macroscale organization protects function when individual components fail.¹⁶ This line of work bridges the gap between the microscale focus of traditional spiking models and the macroscale view of clinical neuroimaging.

Biohybrid Experiments and Closing the Loop with Living Tissue

Perhaps the most striking use of SNNs in brain research is the biohybrid setup, where a neuromorphic chip running a spiking network is physically coupled to living neurons in a dish. In one such arrangement, the activity of a biological neural network in vitro and the activity of a biomimetic hardware network are recorded, processed, and integrated bidirectionally in real time.¹⁷ The hardware network can stimulate the biological network and respond to its output, creating a closed loop that lets researchers test predictions about plasticity, synchronization, and network dynamics in a mixed living-and-artificial system. If a model predicts that a certain pattern of input should trigger long-term strengthening of synapses, the biohybrid setup lets you check whether the biological tissue actually does what the model says it should.

Neuromorphic Hardware and Energy Efficiency

Running SNNs on conventional processors is wasteful in the same way that running a gasoline engine to power a bicycle is wasteful: the hardware was designed for a different kind of computation. Neuromorphic chips, built specifically to exploit the sparse, event-driven nature of spiking networks, can be dramatically more efficient. Benchmarks of existing neuromorphic systems show that for deep network inference, specialized neuromorphic hardware can consume roughly an order of magnitude less energy than a GPU simulation, and certain platforms achieve two orders of magnitude better efficiency than more general-purpose neuromorphic systems like SpiNNaker when running at full utilization.¹⁸

The energy gap matters because brain-scale simulation will require enormous computational budgets. A typical cortical neuron fires only a few spikes per second on average, and between spikes, both the neuron and its outgoing connections are essentially idle. Neuromorphic chips take advantage of this by doing no work when there is no spike to process, unlike GPUs that churn through calculations on every clock cycle regardless. The brain itself appears to have been shaped by similar pressures. Mathematical modeling of action potentials across vertebrate and invertebrate species has shown that many mammalian neurons have biophysical properties that minimize the overlap between inward and outward ionic currents during a spike, reducing energy consumption. The researchers argue that natural selection for energy efficiency helps explain both the shape of the action potential and the underlying ion-channel properties, a finding that offers neuromorphic engineers a biological precedent and an optimization target.¹⁹

Learning Rules and the Training Problem

Training SNNs is harder than training standard deep-learning networks. The core difficulty is that a spike is an all-or-nothing event: a neuron either fires or it does not. That binary behavior means you cannot use the standard calculus-based approach to training, which requires smooth, differentiable functions. The gradient, the signal that tells each connection how to adjust its strength, vanishes at the spike threshold.

Surrogate gradient methods have become the dominant workaround. The idea is to replace the true, discontinuous spike function with a smooth approximation during the training phase only, so that gradients can flow through the network and guide learning. During actual operation, the network still uses real, binary spikes.²⁰ This pragmatic compromise has unlocked the application of deep-learning-style training to spiking networks, making them far more capable on real-world tasks than they were with earlier biologically inspired learning rules alone.

A complementary approach draws on developmental neuroscience. During brain development, the brain first overproduces synapses and then prunes those that are underused, following a “use it or lose it” principle. An adaptive pruning method inspired by this process dynamically removes unnecessary connections during learning without requiring any pre-training or retraining phase. On neuromorphic benchmark datasets, this method produced extremely compressed networks while maintaining strong performance, and it worked for both SNNs and conventional artificial neural networks.²¹

Testing Theories of Consciousness

One of the more speculative applications of SNNs in brain research involves testing theories about conscious experience. Giulio Tononi’s information integration theory proposes that consciousness arises when a system integrates information in a way that is both highly differentiated and highly unified. That theory makes quantitative predictions, but testing them in a living brain is nearly impossible because you would need to measure the information structure of an entire network at once. SNNs offer a tractable alternative. A spiking network of roughly 18,000 neurons and 700,000 connections, equipped with models of emotion and imagination to control the eye movements of a virtual robot, was used to make detailed predictions about where phenomenal states should and should not arise according to the theory.²² No one claims the robot is conscious; the point is that the theory’s predictions can be generated and examined in a system whose internal structure is completely known, which is never the case for a real brain.

The Scaling Bottleneck

For all their biological appeal, SNNs face a stubborn scaling problem. The human brain has roughly 86 billion neurons. Until very recently, the largest SNNs contained fewer than a billion parameters, and most research networks are far smaller still. A recent effort to bridge that gap proposed spiking large language models with 7 to 70 billion parameters, aiming to bring spiking dynamics to the scale where conventional AI already operates.²³ Whether spiking models at that scale will yield the energy and robustness advantages they offer at smaller scales remains an open question, but the attempt itself signals how seriously the field takes the ambition of matching biological complexity.

The bottleneck is not purely computational. Spiking networks have an extra dimension, time, that conventional networks do not. Every spike carries a timestamp, every synapse has a delay, and the network’s behavior can change depending on the precise temporal pattern of input. That richness is exactly what makes SNNs useful for brain research, but it also multiplies the number of parameters that must be set correctly and the amount of data needed to constrain them. For brain modeling specifically, the gap between the circuits we can simulate in detail and the full-scale organ remains vast. The field is progressing by building detailed models of small circuits and linking them, rather than trying to simulate everything at once.