Maximum entropy models of neuronal populations at and off criticality
T. S. A. N. Simões, F. Lombardi, D. Plenz, H. J. Herrmann, L. de Arcangelis · 2025
Read Maximum entropy models of neuronal populations at and off criticality by T. S. A. N. Simões, F. Lombardi, D. Plenz, H. J. Herrmann, L. de Arcangelis (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-14872.txt.
Introduction
T. S. A. N. Simões, F. Lombardi, and three co-authors connect two ways of testing whether brain tissue sits near a critical point. Recordings of neuronal avalanches, cascades of firing cells, often show scaling patterns that suggest the tissue balances between quiet and runaway activity. Separately, maximum entropy models built from the same data carry heat-like signatures of criticality.
The team records spontaneous activity in slices of rat cortex kept alive in the lab, then pushes the tissue toward quiet or excitable states with drugs. Each state is classed as below, at, or above criticality by avalanche behavior. Comparing the two lenses in the same tissue shows how their verdicts line up, a step toward trusting either one as a health marker.
Catalog record
| CREATOR | T. S. A. N. Simões, F. Lombardi, D. Plenz, H. J. Herrmann, L. de Arcangelis |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 64 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-14872
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Maximum entropy models of neuronal populations at and off criticality T. S. A. N. Simões, F. Lombardi, D. Plenz, H. J. Herrmann, L. de Arcangelis Abstract Empirical evidence of scaling behaviors in neuronal avalanches suggests that neuronal populations in the brain operate near criticality. Departure from scaling in neuronal avalanches has been used as a measure of distance to criticality and linked to brain disorders. A distinct line of evidence for brain criticality has come from thermodynamic signatures in maximum entropy (ME) models. Both of these approaches have been widely applied to the analysis of neuronal data. However, the relationship between deviations from avalanche criticality and thermodynamics of ME models of neuronal populations remains poorly understood. To address this question, we study spontaneous activity of organotypic rat cortex slice cultures in physiological and drug-induced hypo- or hyper-excitable conditions, which are classified as critical, subcritical and supercritical based on avalanche dynamics. We find that ME models inferred from critical cultures show signatures of criticality in thermodynamic quantities, e.g. specific heat. However, such signatures are also present, and equally strong, in models inferred from supercritical cultures—despite their altered dynamics and poor functional performance. On the contrary, ME models inferred from subcritical cultures do not show thermodynamic hints of criticality. Importantly, we confirm these results using an interpretable neural network model that can be tuned to and away from avalanche criticality. Our findings indicate that maximum entropy models correctly distinguish subcritical from critical/supercritical systems. However, they may not be able to discriminate between avalanche criticality and supercriticality, although they may still capture a number of important features from neuronal data. I Introduction Biological neural networks need to perform complex functions and continuously adapt. Such abilities rely on cooperative effects among local and distributed neuronal populations, which underlie the emergence of a variety of collective behaviors in the brain. The analogy between collective behaviors of populations of neurons and cooperative phenomena in physical systems undergoing a phase transition suggests that brain networks self-organize to operate at or near criticality 9 , a state that provides several functional advantages 21 ; 49 . This hypothesis is primarily supported by observations of long-range spatio-temporal correlations and neuronal avalanches across species and spatial scales. Neuronal avalanches are cascades of local synchronized activity whose size and duration distributions follow power law behaviors, a hallmark of criticality that imply absence of characteristic scales. First identified in acute slices and longterm slice cultures of rat cortex in vitro 3 , neuronal avalanches have since been reported in vivo in rats 12 , monkeys 38 , and other species 13 ; 39 , including MEG and EEG recordings of the human brain 50 ; 25 ; 44 . Concomitant evidence of tuning to criticality in neuronal systems has come from maximum entropy (ME) models of neural activity, which focused on thermodynamic aspects such as the divergence of the specific heat 54 ; 32 ; 13 ; 26 ; 43 . ME modeling has proven a powerful approach to study the dynamics of biological neural networks 45 ; 55 ; 53 . Spiking activity of neurons can be described as a binary process, \sigma(t) , in which \sigma=\pm 1 represents the state of the neuron at a given time t , i.e. spiking for \sigma=1 and silent for \sigma=-1 40 . This approach allows one to define an empirical distribution of binary activity patterns, P(\bf{\sigma}) , which specifies the probability of observing a given spiking pattern in a population of neurons. For N binary neurons, the distribution P(\bf{\sigma}) for all 2^{N} possible spiking patterns (or states) fully characterizes population dynamics. However, estimating P(\bf{\sigma}) directly from data is often impractical since the number of possible states grows exponentially with the number of neurons 31 . Maximum entropy modeling offers a possible solution to this problem by constructing the least-biased probability distribution that matches selected statistics, such as individual firing rates and pairwise correlations, while maximizing the entropy of P(\bf{\sigma}) . The resulting distribution is mathematically equivalent to the Boltzmann distribution 45 ; 55 ; 11 , which can then be analyzed using thermodynamic tools, such as the fluctuation-dissipation theorem 54 ; 43 ; 51 . This procedure can be interpreted as building a thermodynamics-like framework to describe steady-state properties (e.g., time-averaged firing rates in neural networks) of non-equilibrium systems 14 ; 15 . ME modeling has been extensively applied to neuronal data, such as evoked activity in the salamander retina 45 ; 53 ; 54 ; 17 , the nervous system of the C. Elegans 8 , and in vivo and in vitro neuronal activity from cortical tissue of rodents 43 . In this context, criticality is identified from a maximum in the susceptibility or in the specific heat of the data-inferred model close to unit temperature, corresponding to the observed statistics of neural activity. Such signatures of criticality in neural data can be considered static as opposed to those coming from scaling of neuronal avalanches and long-range correlations, which are intrinsically dynamic. Recently, ME models that take into account temporal dynamics of neural activity have been proposed to enhance (static) evidence of criticality, while matching avalanche statistics 32 . However, to what extent static and dynamic signatures of criticality in neural systems need to coexist and agree in baseline normal condition and how they change when such systems are perturbed away from baseline condition remains poorly understood. Are deviations from dynamical criticality captured by changes in thermodynamic properties of the corresponding maximum entropy models and vice versa? As criticality is increasingly used as a biomarker for brain disorders, these questions become of key relevance also in neuroscience. Here, we address them by analyzing neural cultures in baseline and pharmacologically perturbed conditions, together with an interpretable neural network model whose dynamics can be tuned to criticality 30 ; 51 . We consider three sets of neuronal cultures: one in baseline, physiological condition, one with reduced neural excitability, and one with reduced neural inhibition 48 . Cultures in physiological condition showed neuronal avalanches whose size and duration distributions were consistent with power-laws, and were classified as critical in 48 . In cultures with reduced excitability instead, avalanches were small and short-lived, with exponential size and duration distributions. These cultures were identified as subcritical. Conversely, disinhibited cultures exhibited high excitability and a pronounced increase in the probability of very large avalanches—of the order of the system size—, leading to bimodal-like size/duration distributions. These cultures were classified as supercritical 48 . Importantly, each of these scenarios can be reproduced with our network model by tuning a single parameter 30 . We use a maximum entropy approach to model experimental and numerical data by constraining firing rates, pairwise correlations, and the distribution of synchrony, P(K) , defined as the probability for K sites of the network or, alternatively, K neurons, to be active simultaneously 53 . By studying their thermodynamic properties, we show that the inferred models (mostly) agree with the dynamical classification of hypoexcitable cultures as subcritical, and correctly distinguish them from dynamically critical/supercritical cultures. However, they are not able to discriminate between dynamical criticality and supercriticality. This results are confirmed by the analysis of our integrate-and-fire (IF)neural network, whose inferred ME model closely matches the