Single neuron properties shape chaotic dynamics in random neural networks

Samuel P. Muscinelli, Wulfram Gerstner, Tilo Schwalger

Abstract

Cortical neurons exhibit rich dynamics, shaped by the presence of history-dependent biophysical mechanisms such as adaptation over different time scales. On the other hand, cortical circuits are highly recurrent which can result in chaotic network dynamics even if network nodes have simple dynamics. We analyze how the dynamics of nodes (such as single neurons) and recurrent connections interact to shape the effective dynamics in large randomly connected networks. A novel dynamical mean-field theory for networks of multi-dimensional rate units shows that, despite the nonlinear interactions between units, the recurrent network dynamics inherit, in the chaotic phase, the qualitative features of the linear response properties of the single units. For the case of two-dimensional rate units with strong adaptation, we find that the network exhibits a state of resonant chaos, whose resonance frequency can be predicted from the properties of isolated units. The correlation time of the chaotic network scales with the adaptation timescale of a single unit. Our theory for network analysis can be applied to numerous other mechanisms at the level of single nodes, such as synaptic filtering, refractoriness or spike synchronization. These results advance our understanding of the interaction between single-neuron mechanisms and recurrent connectivity, a fundamental step toward the description of biologically realistic network models in the brain, or, more generally, networks of other physical or man-made complex dynamical units.

Introduction

A fundamental challenge in theoretical neuroscience is bridging the gap between the complex internal biophysics of single neurons and the collective behavior of large-scale neural circuits. Since the late 1980s, the "standard model" for random neural networks has established that large, highly connected networks of simple units naturally exhibit chaotic dynamics. However, these classical models typically utilize "first-order" units—simple switches lacking temporal memory.

In reality, biological neurons are far more sophisticated, exhibiting features like Spike-Frequency Adaptation (SFA) and operating across multiple timescales. This research addresses whether the collective chaos of a network washes out these single-neuron features or if the network's global dynamics "inherit" them. By extending Dynamical Mean-Field Theory (DMFT) to multi-dimensional units, the authors reveal a novel state termed "resonant chaos."

Methodology: Generalized DMFT

The study employs two primary theoretical innovations:

  1. Multi-dimensional DMFT: The authors extend the classic DMFT framework, originally designed for 1D units, to $D$-dimensional units. This allows a massive network of $N$ interacting neurons to be mathematically represented by a single "representative" neuron driven by self-consistent Gaussian noise.
  2. Frequency-Domain Iterative Map: To overcome the difficulty of solving multi-dimensional DMFT in the time domain, the authors developed a semi-analytical method in the frequency domain. By iteratively refining the power spectrum, they can determine the steady-state fluctuations of the network significantly faster than through full-scale microscopic simulations.

The theoretical predictions were validated against numerical simulations of networks containing $N = 1000$ to $2000$ units.

Results: The Inheritance of Single-Neuron Properties

The central finding of this work is the "Inheritance Principle": the power spectrum of a chaotic network is essentially a "sharpened" version of the single neuron's linear response function. If an isolated neuron exhibits a resonance (a preference for specific frequencies), the entire chaotic network will manifest a peak at that same frequency.

Transition from stable to chaotic dynamics in networks of resonant (blue) and non-resonant (red) units.

As shown in Figure 1, the network transitions from a stable fixed point to chaos at a critical coupling strength $g_c$. In the chaotic regime ($g > g_c$), networks composed of resonant units (blue) enter a state of resonant chaos. Unlike the broad-band noise seen in standard chaotic models (red), resonant chaos is characterized by robust, narrow-band stochastic oscillations—a "noisy rhythm" emerging from the interplay of single-unit adaptation and recurrent connectivity.

Key Findings:

  • Stabilization through Adaptation: Adaptation mechanisms stabilize the network, requiring a higher critical coupling strength $g_c$ to trigger chaos compared to non-adaptive models.
  • Timescale Scaling: The correlation time of the network—the duration over which it retains memory of its state—scales directly with the adaptation timescale of the individual units.
  • Chaos as a Signal Booster: In the chaotic regime, adaptation suppresses low-frequency chaotic "noise" more effectively than it suppresses external signals. This leads to a significant enhancement of the Signal-to-Noise Ratio (SNR) for slow, behaviorally relevant stimuli, a phenomenon not observed in simpler 1D networks.

Conclusion and Future Directions

This research demonstrates that single-neuron properties like adaptation do not merely dampen activity but fundamentally shape the rhythmic structure of network-wide chaos. While this study utilizes rate models, which describe average activity, the authors identify the extension to spiking neuron models as a critical next step. Furthermore, while the current work assumes random connectivity, future investigations could explore how resonant chaos interacts with the highly structured architectures (motifs and layers) found in biological brains. The frequency-domain framework developed here is particularly well-suited for future studies involving power-law adaptation across vast ranges of timescales.

Summary updated

Abstract

Cortical neurons exhibit rich dynamics, shaped by the presence of history-dependent biophysical mechanisms such as adaptation over different time scales. On the other hand, cortical circuits are highly recurrent which can result in chaotic network dynamics even if network nodes have simple dynamics. We analyze how the dynamics of nodes (such as single neurons) and recurrent connections interact to shape the effective dynamics in large randomly connected networks. A novel dynamical mean-field theory for networks of multi-dimensional rate units shows that, despite the nonlinear interactions between units, the recurrent network dynamics inherit, in the chaotic phase, the qualitative features of the linear response properties of the single units. For the case of two-dimensional rate units with strong adaptation, we find that the network exhibits a state of resonant chaos, whose resonance frequency can be predicted from the properties of isolated units. The correlation time of the chaotic network scales with the adaptation timescale of a single unit. Our theory for network analysis can be applied to numerous other mechanisms at the level of single nodes, such as synaptic filtering, refractoriness or spike synchronization. These results advance our understanding of the interaction between single-neuron mechanisms and recurrent connectivity, a fundamental step toward the description of biologically realistic network models in the brain, or, more generally, networks of other physical or man-made complex dynamical units.
BibTeX
@article{muscinelli2019single,
  author = {Samuel P. Muscinelli and Wulfram Gerstner and Tilo Schwalger},
  title = {{Single neuron properties shape chaotic dynamics in random neural networks}},
  journal = {PLOS Computational Biology},
  year = {2019},
  doi = {10.1371/journal.pcbi.1007122},
}
Figure