Neurons are a motley crew, varying widely in connectivity and firing properties. But computational models of the brain usually treat neurons as identical.
“The convention in the field is homogeneity,” says Scott Rich, assistant professor at the University of Connecticut. Heterogeneity makes models mathematically complex and unwieldy, so researchers have shied away from incorporating it, he says. “But it shouldn’t be [omitted] if we really want to understand how the brain works.”
Heterogeneity has beneficial effects on neural computations, several new lines of evidence suggest. It can improve decision making by increasing a neural network’s flexibility, according to three published studies, as well as unpublished findings presented at a workshop on neuronal heterogeneity at this year’s Computational Neuroscience Meeting in Halifax, Nova Scotia. Too much diversity in firing properties can destabilize neural networks, the work confirms. But none can lead to pathological states: In humans, loss of heterogeneity in pyramidal neurons’ intrinsic firing properties may cause sudden, pathological shifts into synchronized firing. Homogenization could itself trigger the transition into seizure-like, hypersynchronous dynamics, according to work analyzing tissue from people with epilepsy.
And yet the right ‘Goldilocks’ amount can help a network switch flexibly between distinct memories or representations.
Just as theoretical neuroscientists are coming to appreciate the benefits of neuronal heterogeneity, an improved mathematical toolkit is making it easier for them to incorporate it into neural network models, says Richard Gast, a postdoctoral researcher in Ann Kennedy’s lab at The Scripps Research Institute who co-organized the workshop with Rich and presented his work there. The explosion of sequencing and tracing technologies has highlighted vast cell type diversity in the brain; even within a single, narrowly-defined cell type, no two neurons are the same, he says.
“Heterogeneity shouldn’t be a bug, it should be a feature,” says Martin Nawrot, professor of computational systems neuroscience at the University of Cologne in Germany, who was not involved in any of the work presented at the workshop. Incorporating heterogeneity into future models is necessary for understanding the brain’s unique computations, he adds.
A
ttractor networks are popular models for memory systems because their stable states persist even after an input disappears, or individual neurons change their own activity. One way to picture the computational energy of this network is as a hilly landscape full of valleys, in which each valley represents a stored memory. Heterogeneity flattens this landscape, which is one reason modelers avoid it, Gast says.Networks with neurons that differ in how easily they fire have shallower valleys than homogeneous ones. They also hold on to fewer distinct memories, but can move from one to another more flexibly. Their variability makes the network’s activity richer and easier for downstream circuits to read out, improving its performance on decoding tasks, Gast and his colleagues published in PNAS.
That finding was purely computational, but applying the same principles to mouse cortical neuron recordings yields similar results. The researchers used firing threshold measurements of the real cortical neurons to build network models. Compared with homogeneous neurons, cortical neurons whose properties vary widely support fewer distinct stable states but need less input to snap into a new state, according to a preprint posted in July and presented at the workshop.
The same trade-off may shape how animals perform tasks, according to a study published in December 2025 in Neuron. The relationship between auditory task performance and arousal famously follows an inverted-U shape: If an animal (or a human, for that matter) is not attentive enough or too keyed up, their performance suffers. The mechanisms behind this well-studied phenomenon have been “a mystery for 100 years,” says study investigator Luca Mazzucato, associate professor at the University of Oregon.
Mazzucato and his colleagues measured how well they could predict which tone a mouse heard by correlating its pupil size—which tracks its level of arousal—with the activity of neuronal populations in the auditory cortex.
In a network model of the circuit, too little heterogeneity made it difficult for the system to switch between states, and too much pushed it into chaos. But an intermediate amount allowed the network to switch cleanly between states, sharpening its ability to tell low- and high-frequency tones apart. And neuromodulators may be the dial that tunes that heterogeneity up and down. Arousal is linked to the brain’s norepinephrine levels, which can tune how heterogeneous neurons’ excitability is.
H
eterogeneity in a neuron’s connectivity degree—how many neurons connect to it—is another feature that can shape network dynamics, according to a preprint published in March by Mazzucato and his colleagues and presented at the Computational Neuroscience Meeting. Cortical neurons vary enormously in how many other neurons they connect to, the team found using mouse connectomics data. Popular neurons have a lot of connections, whereas more solitary neurons have fewer. These variations alter how quickly neurons return to a baseline state, the study showed.That mix of timescales lets a single input echo through a network at different speeds, Nawrot says, so some neurons react instantly while others hold on to information and integrate it over time.
Different types of heterogeneity can also work in tandem, according to a June 2025 study involving a Hopfield network, an attractor network in which every neuron is connected to all other neurons. The researchers tweaked the network so that not all neurons connected and the cells varied in how many different stimuli they respond to, a property known as coding level.
Adding either kind of heterogeneity alone hurt the network’s memory capacity, but combining the two restored it. When popular neurons have high coding levels, the network works roughly as well as a fully homogeneous network. “It’s not just, ‘heterogeneity is good’ or ‘heterogeneity is bad,’ but it depends on the computation and how this heterogeneity is structured,” says study investigator Gaia Tavoni, assistant professor of neuroscience at Washington University in St. Louis.
Some heterogeneity “always makes things better,” Nawrot says. Existing attractor models struggle to reproduce the trial-to-trial variability seen in real neural and behavioral data without becoming unstable—a problem that can be fixed, at least in some instances, by clustering excitatory and inhibitory neurons, Nawrot says. Without heterogeneity, networks are prone to falling into synchrony and getting stuck in a rhythm.
Heterogeneity in neurons’ firing properties may also prevent the kind of runaway synchrony that produces an idealized rhythm in a computational model, instead enabling shorter, weaker and more biologically realistic oscillations to occur, according to unpublished data that Rich shared at the workshop.
The apparent contradiction—heterogeneity eroding stability in some studies, while reinforcing it in others—comes down to what a specific brain region needs to do, Rich says: “The brain is complex enough that both answers are probably right, in different scenarios and for different functions.” Cells in the thalamus, for example, are largely homogeneous. But because the thalamus acts as an on-off gate, it may not need the same sort of variability as cortical regions.
Future models should aim to better capture the heterogeneity seen in biological systems, he adds. “Even the studies looking at heterogeneity, the heterogeneity is still kind of implemented somewhat ad hoc,” rather than tied to real, measured biology, he says. “That’s the next step.”
