Dynamically rich states in balanced networks induced by single-neuron dynamics

Moritz Drangmeister, Rainer Engelken, Jan‐Hendrik Schleimer, Susanne SchreiberbioRxiv 2025 · DOI

TL;DR

This paper reveals that the way individual neurons initiate action potentials (their "excitability class") can profoundly influence the complex dynamics of large neural networks, even with weak synaptic connections. By switching a neuron's dynamics from a standard "saddle-node on invariant circle" (SNIC) to a "saddle homoclinic orbit" (HOM) bifurcation, researchers observed significantly more variable and "bursty" network activity, improved information encoding at higher frequencies, and enhanced network chaos. This highlights the crucial role of single-neuron biophysics in shaping network-level computation.

What Problem Does This Paper Address?

The brain's ability to process complex information is thought to rely on neural networks exhibiting rich, dynamic activity patterns with highly variable firing rates. Traditionally, explanations for these complex states have focused on network-level properties like strong synaptic connections or slow synaptic dynamics. However, the intrinsic biophysical properties of individual neurons, specifically how they generate action potentials, have often been considered less critical and potentially "washed out" in large, balanced networks. This paper challenges that assumption by investigating whether a fundamental change in a single neuron's firing mechanism can, in fact, drive complex network-level dynamics.

Methods

The researchers used a simplified mathematical model of neurons called the quadratic integrate-and-fire (QIF) neuron. The key to their approach was manipulating the "reset voltage" of these neurons. This parameter controls the type of mathematical bifurcation that governs how a neuron initiates an action potential. By changing the reset voltage, they could switch the neuron's dynamics between two classes:

  • SNIC (Saddle-Node on Invariant Circle): This is a common type of bifurcation where firing rate increases gradually with input.
  • HOM (Saddle Homoclinic Orbit): This bifurcation leads to a more complex behavior where neurons can exhibit bistability, meaning they can either rest or fire in a limit cycle. This bistability, especially in the presence of noise, can lead to "bursty" firing patterns (periods of rapid firing interspersed with silence).

They then simulated large, randomly connected, balanced networks (where excitation and inhibition are roughly balanced) composed of these QIF neurons. They analyzed network activity by examining:

  • Spike Raster Plots: Visualizing when each neuron fires over time.
  • Inter-Spike Interval (ISI) Distributions: Analyzing the timing between individual spikes.
  • Power Spectra: Examining the frequency content of neural activity.
  • Coefficient of Variation (CV): Quantifying the variability of spike timings.
  • Lyapunov Exponents: Measuring the degree of chaos in the network dynamics.
  • Mutual Information Rate (MIR): Quantifying how much information the network can encode about an input signal.

They also explored heterogeneous networks, where a mix of SNIC and HOM neurons were present, and examined excitatory-inhibitory (EI) balanced networks to assess the generality of their findings.

Datasets

Not discussed in the paper. The study relies on simulated neural networks rather than experimental datasets.

Key Insights and Ideas

  • Single-Neuron Biophysics Matters: The most significant insight is that the fundamental way a neuron initiates a spike (its excitability class) is not "washed out" in large networks. Instead, it can be a primary driver of complex network dynamics.
  • HOM Bifurcation Drives Burstiness: Switching to HOM dynamics, even with weak synaptic coupling, reliably induces super-Poissonian spiking variability and "burst-like" firing patterns in balanced networks. This is a departure from explanations that rely solely on strong synaptic interactions.
  • Shift in Information Encoding: HOM neurons shift information encoding towards higher frequency bands. This suggests that the specific biophysical properties of neurons can tune a network's sensitivity to different temporal scales of input.
  • Enhanced Network Chaos: The HOM regime amplifies network chaos, but in a structured way, leading to a lower-dimensional but more chaotic manifold. This "edge of chaos" regime is theorized to be optimal for computation.
  • Heterogeneity Amplifies Effects: Even a small fraction of HOM neurons in a network can significantly enhance chaos and variability, suggesting that biological heterogeneity might play a crucial role in neural computation.

Key Results

  • Networks of HOM neurons exhibit significantly more irregular spiking (higher CV) compared to SNIC networks, characterized by long silences punctuated by bursts of spikes.
  • The power spectrum of HOM neuron activity shows increased power at lower frequencies, reflecting the slow switching between resting and bursting states.
  • Information encoding shifts to higher frequencies in HOM networks, with a band-pass filtering effect becoming more pronounced in network simulations.
  • Lyapunov exponents increase in the HOM regime, indicating heightened chaotic dynamics. The Kolmogorov-Sinai entropy rate also peaks in this regime.
  • In heterogeneous networks, even ~25% HOM neurons can substantially increase network chaos and variability, with the largest Lyapunov exponent showing a significant jump.
  • These effects persist in mixed excitatory-inhibitory balanced networks.

Limitations and Future Work

  • Model Simplification: The study uses a simplified QIF neuron model with pulse coupling. While this allows for analytical tractability and efficient simulation, more complex conductance-based neuron models and dynamic synapses might yield nuanced differences.
  • Biological Parameter Translation: The paper suggests that biological factors like temperature or ion channel changes can induce HOM dynamics. However, the precise translation of these factors into the QIF model's reset voltage parameter warrants further investigation in more detailed models.
  • Future Directions: The authors suggest exploring HOM dynamics in task-specific spiking networks, leveraging the enhanced dynamical repertoire for optimal computational performance. They also propose that neuromorphic hardware could be used to implement and dynamically control these excitability classes, potentially leading to more efficient and biologically realistic artificial neural networks. The findings also open avenues for experimental validation by comparing network behavior with varying proportions of neurons exhibiting HOM-like dynamics.

Summary updated

Abstract

Network states with rich dynamics and highly variable firing rates of individual neurons are prominent in experimental observations and thought to benefit complex information processing and learning. Such states have been proposed to arise from properties of network coupling, like a strong connectivity or slow synaptic dynamics. Here, we identify an alternative mechanism based on weak synaptic coupling and intrinsic cellular dynamics. We show that a switch in the cellular excitability class of action-potential generation (via a switch in the underlying mathematical bifurcation), further amplified by recurrent interactions, results in super-Poissonian spiking variability in random balanced networks. Information encoding is shifted to higher frequency bands and collective chaos in the network is enhanced when intrinsic cellular dynamics follow a saddle homoclinic orbit (HOM) bifurcation. The robust effect links the biophysics of individual neurons to collective dynamics of large random networks, highlighting the relevance of single-cell dynamics for computation in physiological and artificial networks.
BibTeX
@misc{drangmeister2025dynamically,
  author = {Moritz Drangmeister and Rainer Engelken and Jan‐Hendrik Schleimer and Susanne Schreiber},
  title = {{Dynamically rich states in balanced networks induced by single-neuron dynamics}},
  year = {2025},
  doi = {10.1101/2025.02.28.640576},
  howpublished = {bioRxiv},
}
Figure