Collective Dynamics of Excitation and Inhibition: An Ising Model Approach

Companion post to:
Ensemble Inhibition and Excitation in the Human Cortex: An Ising-Model Analysis with Uncertainties
Cristian Zanoci, Nima Dehghani, and Max Tegmark Physical Review E . Phys. Rev. E 99, 032408 DOI: https://doi.org/10.1103/PhysRevE.99.032408

Overview

One of the recurring ambitions of theoretical neuroscience is to find an effective macroscopic description of neural population activity: a description that is not merely a catalog of individual spikes, but also not so coarse that it erases the collective structure of the circuit. This problem is naturally close to statistical physics. A cortical population is a many-body system: many interacting units, strong correlations, partial observations, limited samples, and emergent collective states. The central question is what kind of reduced model can capture the relevant structure of the activity without pretending to model the full microscopic machinery of the brain.

In this paper, we studied this question using pairwise maximum-entropy models, also known as Ising models, applied to human cortical spiking data. The data consisted of multielectrode recordings from layers II/III of the human temporal cortex across the wake–sleep cycle. The neurons had been classified as putative excitatory or inhibitory, allowing us to ask not only whether an Ising model can capture collective activity, but also how the balance between excitation and inhibition shapes the collective statistics.

The paper had two main goals. The first was methodological: to develop a statistically more careful approach to fitting neural Ising models, including uncertainty estimates on the inferred parameters. The second was biological and physical: to use this approach to examine how excitatory and inhibitory ensembles contribute to cortical collective behavior across wakefulness, light sleep, and deep sleep.

The central result is that the pairwise model captures a large fraction of the correlative structure in the data, but only when excitation and inhibition are modeled together. If inhibitory neurons are omitted, the synchrony among excitatory neurons is dramatically overestimated. This result is important because it shows that the apparent collective behavior of a neuronal subpopulation can be badly mischaracterized when the balancing subpopulation is not explicitly included.

At the same time, the thermodynamic analysis shows signatures reminiscent of criticality, including peaks in heat capacity near the fitted operating point. But the paper argues for caution: such signatures may reflect long-range correlations, shared inputs, subsampling, or properties of the inference procedure rather than true critical dynamics in the biological circuit.

From cortical spiking to an Ising model

The starting point is a population of (N) neurons. The spike train of each neuron is discretized into time bins of width (\Delta t). In each bin, neuron (i) is assigned a binary state,

\[\sigma_i = \begin{cases} +1, & \text{if neuron } i \text{ spikes in the bin}, \\ -1, & \text{otherwise}. \end{cases}\]

Thus, at each time bin the population is represented by a binary vector

\[\sigma = (\sigma_1,\ldots,\sigma_N) \in \{-1,+1\}^N.\]

The empirical data provide the mean activity of each neuron,

\[m_i = \langle \sigma_i \rangle,\]

and the pairwise second moments,

\[Q_{ij} = \langle \sigma_i \sigma_j \rangle.\]

The maximum-entropy principle then asks for the least structured probability distribution over population activity patterns that matches these empirical first and second moments. The result is the pairwise maximum-entropy model,

\[P(\sigma) = \frac{1}{Z} \exp[-H(\sigma)],\]

with Hamiltonian

\[H(\sigma) = -\sum_i h_i \sigma_i -\sum_{i,j} J_{ij} \sigma_i \sigma_j .\]

Here, (h_i) is a bias term describing the intrinsic tendency of neuron (i) to be active or silent, and (J_{ij}) is an effective pairwise coupling between neurons (i) and (j). Positive (J_{ij}) favors co-activation or co-silence, while negative (J_{ij}) disfavors matched states. The partition function,

\[Z = \sum_{\sigma} \exp[-H(\sigma)],\]

normalizes the distribution.

This is formally the Ising model, but with an important difference from the canonical lattice systems of statistical physics. There is no regular lattice, no translational symmetry, and no simple nearest-neighbor coupling. The inferred coupling matrix is dense and heterogeneous. In this sense, the model is closer to a fully connected spin glass than to a ferromagnet on a regular grid.

For neuroscience, this is both a strength and a limitation. It is a strength because the model can capture heterogeneous correlation structure without imposing an artificial geometry. It is a limitation because the inferred couplings are effective statistical interactions, not direct synaptic connections. A negative (J_{ij}), for example, should not be naively interpreted as a monosynaptic inhibitory connection. It is a parameter in an effective equilibrium model fit to binned spiking activity.

Why uncertainty matters

A major motivation for the paper was that the neural Ising-model literature had produced partly conflicting conclusions. Some studies found that pairwise maximum-entropy models captured neural population activity surprisingly well. Others argued that pairwise models fail, especially for local populations or for systems with strong higher-order structure.

Part of the difficulty is methodological. Fitting an Ising model is an inverse problem. Given empirical moments (m_i) and (Q_{ij}), we must infer the parameters (h_i) and (J_{ij}). For (N) neurons, the number of parameters is (N(N+1)/2), and the number of possible activity patterns is (2^N). Exact evaluation of the partition function becomes intractable for large (N). This makes model fitting computationally difficult and also makes uncertainty quantification nontrivial.

In many applications, one reports a fitted coupling matrix as if it were a single determinate object. But in finite data, especially with sparse spiking and subsampling, some inferred parameters may be stable while others are not. Without uncertainty estimates, it becomes difficult to know which biological or physical conclusions are reliable.

The paper therefore introduced an uncertainty-estimation procedure based on adaptive Markov-chain Monte Carlo in parameter space. After the main optimization algorithm finds a maximum-likelihood solution to the inverse Ising problem, the method performs a random walk around that solution in the high-dimensional parameter space. This allows one to estimate uncertainties on both the bias parameters (h_i) and the couplings (J_{ij}).

This is conceptually important. The model is not only a fit; it becomes an object with error bars. That makes it possible to ask which parameters, and which derived quantities, are statistically meaningful.

Reliable neurons and the problem of finite data

The uncertainty analysis also exposed a practical issue: not all neurons are equally reliable for Ising-model inference.

When the optimization was run multiple times with different initial conditions, the inferred parameters were not always identical. This could have two explanations. One possibility is that the underlying inverse problem genuinely has multiple relevant optima. Another is that finite sampling noise pushes the empirical moments away from their true values, generating instability in the inferred parameters.

The paper addressed this by asking how reproducible the inferred parameters are when only the most active neurons are retained. The result was clear: parameters associated with more active neurons are much more reliably inferred. This is expected statistically. If a neuron rarely spikes, then its firing rate and correlations are estimated from relatively few events, and the inferred parameters involving that neuron become noisy.

To quantify reliability, the paper used cosine similarity between parameter vectors inferred from different runs. Neurons were retained when the inferred parameters were stable enough across initializations. In practice, this selected approximately the neurons firing in at least about (5\%) of the time windows, leaving a reliable subpopulation for subsequent analysis.

This step is more than a technical filter. It reflects a broader point about modeling neural populations. The question is not simply whether one can write down an Ising model for all recorded neurons. The question is whether the available data support reliable inference for the degrees of freedom included in the model. In this sense, the paper treats model fitting as a statistical physics problem under experimental constraints, rather than as a purely formal exercise.

Pairwise models capture much of the collective structure

The first major result is that the Ising model strongly outperforms the independent model.

The independent model matches only the firing rates. It assumes

\[P(\sigma) = \prod_i P_i(\sigma_i),\]

and therefore ignores interactions among neurons. This model provides a useful baseline: if neural population activity were mostly explainable by individual firing rates alone, then the independent model would already predict the observed population patterns.

It does not. The independent model strongly underpredicts synchronous events, especially activity patterns in which many neurons spike within the same time bin. This is a familiar failure mode: independent neurons do not produce enough coordinated activity.

The Ising model, by contrast, captures much more of the observed population structure. It predicts the frequencies of activity patterns far better than the independent model and gives a much better account of the distribution of synchronous spiking events.

Information-theoretically, this improvement can be quantified using entropy. Let

\[S_1\]

be the entropy of the independent model,

\[S_2\]

the entropy of the pairwise Ising model, and

\[S_N\]

the entropy of the empirical distribution. The total multi-information is

\[I_N = S_1 - S_N,\]

which measures how much statistical structure is present beyond independent firing. The pairwise contribution is

\[I_2 = S_1 - S_2.\]

The ratio

\[\frac{I_2}{I_N}\]

then measures the fraction of the total correlation structure captured by the pairwise model.

Across sleep states and neuron types, the Ising model captured approximately (80\%)–(95\%) of the correlation structure. This is not a claim that the cortex is literally a pairwise equilibrium Ising system. Rather, it shows that a large fraction of the collective structure in these binned cortical spike patterns is captured by first- and second-order constraints.

That result is notable because the recordings sample only a tiny fraction of the neurons in the local cortical circuit, and the model ignores the vast number of unobserved neurons and inputs. The success of the pairwise model therefore suggests either that the recorded neurons carry substantial information about the relevant low-dimensional collective state, or that pairwise correlations are strong enough statistical summaries to capture much of the observed population structure.

Excitation and inhibition must be modeled together

The most biologically important result concerns excitation and inhibition.

The neurons in the data were classified as putative excitatory or inhibitory based on spike waveform features and cross-correlogram evidence. This allowed us to fit models to mixed E/I populations and to E-only or I-only subpopulations.

When both excitatory and inhibitory neurons were included, the Ising model performed well. But when excitatory neurons were modeled alone, the model dramatically overpredicted synchrony among them. In other words, if one observes only the excitatory subpopulation and tries to explain its collective activity without explicitly modeling inhibition, the inferred model allows too much coordinated excitation.

This is a key point. The cortex is not merely a network of excitatory units with some generic regularization. Inhibition is not a minor correction. It is part of the dynamical organization that prevents runaway synchrony and shapes the collective vocabulary of the circuit.

The coupling matrix reflected this structure. Couplings among inhibitory neurons and among excitatory neurons were mostly positive. Couplings between inhibitory and excitatory neurons were mixed, with both positive and negative values. This is consistent with the fact that the E/I relationship is not a simple one-dimensional opposition. Inhibitory neurons can be driven by excitatory activity, can coordinate with one another, and can suppress or gate excitatory ensembles. The inferred couplings are effective interactions, but their pattern points to the importance of modeling both subpopulations simultaneously.

The asymmetry between E-only and I-only models is also informative. Inhibitory neurons could be modeled reasonably well on their own, though ignoring excitation led to a slight underestimation of their synchrony. Excitatory neurons, by contrast, were badly mischaracterized when inhibition was absent. This suggests that the excitatory population is especially dependent on the balancing and sculpting effects of inhibition.

For systems neuroscience, this reinforces the view that E/I balance should not be treated only as an average equality between excitation and inhibition. The relevant object is the ensemble structure: how excitatory and inhibitory populations jointly shape the distribution of possible cortical activity patterns.

Sleep state dependence

The recordings spanned wakefulness, light sleep, and deep sleep. This allowed the paper to ask whether the collective statistics of excitatory and inhibitory neurons change across behavioral state.

Firing rates are already known to vary strongly across sleep states. The paper showed that the second-order structure also changes. In particular, inhibitory–inhibitory correlations were higher during deep sleep than during wakefulness. This is consistent with the large-scale coherent slow oscillations characteristic of slow-wave sleep.

This result connects naturally to prior work showing that inhibitory neurons can play a central role in regulating cortical oscillations and state-dependent dynamics. During deep sleep, the cortex enters a regime with stronger slow fluctuations and more coherent population-level activity. The increased I-I correlation observed here suggests that inhibitory ensembles participate strongly in this state-dependent reorganization.

At the same time, the Ising model continued to capture a large fraction of the pairwise structure across states. This is interesting because it suggests that the form of the effective model remains useful even as the underlying cortical state changes. The parameters change, the correlation structure changes, but the pairwise maximum-entropy framework remains a meaningful reduced description of the collective activity.

Thermodynamics, heat capacity, and criticality

Because the fitted model has the form of a Boltzmann distribution, one can formally introduce a temperature-like scaling parameter. If the fitted model corresponds to (T=1), then varying (T) defines a family of models,

\[P_T(\sigma) = \frac{1}{Z(T)} \exp[-H(\sigma)/T].\]

This does not mean that the biological cortex is literally being heated or cooled. The temperature is a model parameter that rescales the energy landscape. But this construction allows one to compute thermodynamic quantities such as entropy and heat capacity.

The heat capacity can be written as

\[C(T) = \frac{\langle E^2 \rangle - \langle E \rangle^2}{T^2},\]

where (E = H(\sigma)). In this setting, heat capacity measures the variance of surprise. A peak in heat capacity means that the system has large fluctuations in the probability of activity patterns. In physical systems, diverging heat capacity is associated with criticality and phase transitions.

The paper found that the heat capacity peaks near (T=1), the operating point of the fitted model, and that the entropy changes rapidly near the same point. These features are reminiscent of critical behavior. In a critical regime, a system can exhibit large susceptibility, long-range correlations, and a rich repertoire of accessible states. Such ideas have long been attractive in neuroscience because they suggest a possible explanation for how cortical networks balance stability and flexibility.

However, the paper takes a deliberately cautious position. Signatures of criticality in fitted maximum-entropy models do not necessarily prove that the biological system is truly critical. There are several reasons.

First, shared inputs and latent variables can induce correlations among recorded neurons. Second, multielectrode recordings dramatically subsample the underlying circuit. Third, long-range correlations in the data can force an Ising model toward a near-critical parameter regime simply because that is how the model represents such correlations. Fourth, previous analyses of related data suggested multi-exponential structure rather than clean power-law scaling.

Thus, the thermodynamic analysis is valuable, but its interpretation must be careful. The right conclusion is not simply “the cortex is critical.” A more precise conclusion is that the fitted pairwise model has thermodynamic signatures near criticality, and these signatures may reflect long-range correlation structure in the data. Whether that correlation structure corresponds to true biological criticality, asynchronous irregular dynamics, shared input, subsampling effects, or some mixture of these remains a deeper question.

This distinction matters. Statistical physics gives neuroscience powerful tools, but the mapping between a fitted statistical-mechanical model and the biological circuit is not automatic. The model reveals structure; it does not by itself settle mechanism.

Why this matters for complex systems

From the perspective of complex systems, this paper is an example of how one can move between microscopic data and macroscopic organization.

The microscopic data are spikes from individual neurons. The macroscopic object is the probability distribution over collective activity patterns. The Ising model sits between these levels. It does not model every ion channel, synapse, or cellular process. Instead, it asks whether the collective activity can be approximated by an effective energy landscape constrained by firing rates and pairwise correlations.

This is precisely the kind of move that statistical physics often makes. One identifies the relevant degrees of freedom, writes down a reduced model, fits or derives effective interactions, and studies the resulting collective behavior.

But biological data impose complications that are less common in idealized physical systems. The system is strongly subsampled. The units are heterogeneous. The observations are finite. The data are nonstationary across behavioral states. The inferred interactions are effective rather than microscopic. These complications do not make the statistical-physics approach invalid; they make uncertainty quantification and cautious interpretation essential.

The paper therefore contributes not only a result about cortical E/I dynamics, but also a methodological stance: effective theories of neural population activity should come with uncertainty estimates, reliability checks, and explicit attention to what the model can and cannot mean biologically.

Relevance for AI and machine learning

Although the paper is about human cortical recordings, its implications extend naturally to artificial intelligence, especially energy-based models, recurrent networks, spiking neural networks, and biologically inspired architectures.

The Ising model is closely related to classical energy-based models such as Boltzmann machines and Hopfield networks. In all of these systems, computation can be understood as movement through an energy landscape. Patterns are not processed independently; they are shaped by interactions among units. This makes the paper relevant to AI in at least three ways.

First, the result emphasizes the computational importance of inhibition. Modern artificial neural networks are often built from units that do not have an explicit excitatory or inhibitory identity. In biological circuits, however, inhibitory neurons are not merely negative weights. They form structured subpopulations that regulate synchrony, gate activity, shape oscillations, and stabilize dynamics. The finding that omitting inhibitory neurons causes a dramatic overestimation of excitatory synchrony is a biological warning against architectures that lack explicit balancing mechanisms.

For AI, this suggests that E/I-like structure may be useful not only as a biological detail, but as an inductive bias. Networks with explicit inhibitory populations, balancing constraints, or structured negative feedback may be better able to control runaway activation, maintain sparse representations, and regulate transitions between stable and flexible dynamical regimes.

Second, the paper highlights the importance of uncertainty in high-dimensional model inference. Many AI systems are trained in high-dimensional parameter spaces with complicated loss landscapes. Yet we often report single trained models without a clear account of which learned parameters or structures are stable. The adaptive-MCMC uncertainty analysis in this paper is not directly scalable to today’s largest AI models, but the principle is relevant: when interpreting a fitted model, especially one meant to reveal structure, stability and uncertainty are part of the result.

Third, the thermodynamic analysis connects to a recurring theme in AI: the role of temperature, entropy, and critical-like regimes in generative systems. In modern generative models, temperature-like parameters often regulate the tradeoff between stereotyped and diverse outputs. In recurrent and energy-based models, the shape of the energy landscape determines whether the system settles into a small number of attractors or explores a richer repertoire of states. Biological cortical networks may offer clues about how to tune systems near regimes that balance robustness with flexibility.

The lesson is not that AI should copy the cortex literally. Rather, cortical circuits provide examples of dynamical architectures in which excitation, inhibition, stochasticity, and collective constraints jointly shape computation. For those interested in physical computation and bio-inspired AI, this is a useful bridge: intelligence may depend not only on feedforward function approximation, but also on the controlled organization of state space.

What the paper does not claim

It is important to state what this paper does not claim.

It does not claim that the cortex is literally an equilibrium Ising magnet. The Ising model here is an effective statistical model of binned spiking activity.

It does not claim that the inferred couplings are direct synaptic weights. They are effective pairwise interactions that summarize statistical dependencies among recorded neurons.

It does not claim that all neural correlations are pairwise. Higher-order interactions, temporal dependencies, common input, and latent variables are all important and may become essential in other regimes or at other scales.

It does not claim to prove cortical criticality. The thermodynamic signatures are suggestive, but the interpretation is explicitly cautious.

The value of the paper lies instead in showing that a statistically disciplined pairwise model can capture much of the collective structure of human cortical spiking, that uncertainty estimation is necessary for interpreting the inferred parameters, and that excitation and inhibition must be modeled jointly to avoid misleading conclusions about synchrony.

Outlook

Several natural extensions follow from this work.

One direction is temporal. The model in the paper uses equal-time correlations within time bins. A spatiotemporal maximum-entropy model could include correlations across time lags, allowing one to ask which neurons tend to precede or follow others and how excitatory and inhibitory interactions unfold dynamically.

Another direction is higher-order structure. Pairwise models capture a large fraction of the observed correlations, but not all of them. Adding higher-order interactions may improve the fit, especially in regimes where population-wide events, nonlinear input integration, or common drive dominate. The challenge is that higher-order models require many more parameters and much more data.

A third direction is latent-variable modeling. Because any multielectrode recording samples only a tiny fraction of the local circuit, unobserved neurons and shared inputs can strongly shape the observed statistics. Combining maximum-entropy models with latent-variable approaches may help separate direct pairwise structure from common drive and hidden state fluctuations.

Finally, this line of work can inform artificial systems. If biological intelligence relies in part on structured dynamical balance, then AI architectures may benefit from more explicit mechanisms for controlling synchrony, maintaining stable-but-flexible collective states, and representing uncertainty in learned interactions.

Summary

This paper used an uncertainty-aware Ising-model framework to study excitatory and inhibitory ensembles in the human temporal cortex across the wake–sleep cycle. The main findings are:

Pairwise maximum-entropy models capture a large fraction of the collective structure in cortical spiking activity.
The model strongly outperforms an independent-neuron baseline.
Reliable inference requires attention to finite-data effects and parameter uncertainty.
Excitatory and inhibitory neurons must be modeled together; omitting inhibition leads to a dramatic overestimation of excitatory synchrony.
Inhibitory–inhibitory correlations increase during deep sleep.
Thermodynamic quantities show criticality-like signatures, but these should be interpreted cautiously because subsampling, shared input, and inference effects can produce similar signatures.
The results have implications for statistical physics, systems neuroscience, and biologically inspired AI, especially for energy-based models, recurrent dynamics, and E/I-balanced architectures.

The broader message is that cortical activity can be studied as a many-body system, but doing so responsibly requires both physical modeling and statistical humility. The Ising model is not the cortex. But when fit carefully, with uncertainty estimates and biological constraints, it becomes a useful lens for seeing how excitation, inhibition, and collective structure shape the neural state space.