[Paper Review] Field theory for biophysical neural networks

Si-Wei Qiu, Carson C. Chow|arXiv (Cornell University)|Nov 5, 2014

Neural dynamics and brain function7 references3 citations

TL;DR

This paper develops a field-theoretic framework for modeling biophysically realistic neural networks using quantum field theory techniques, enabling systematic analysis of finite-size effects through path integrals and effective actions. It derives mean field and correlation-corrected equations (activity equations) that capture higher-order statistical moments, offering a non-perturbative approach to study neural network dynamics beyond the infinite-N limit.

ABSTRACT

The human brain is a complex system composed of a network of hundreds of billions of discrete neurons that are coupled through time dependent synapses. Simulating the entire brain is a daunting challenge. Here, we show how ideas from quantum field theory can be used to construct an effective reduced theory, which may be analyzed with lattice computations. We give some examples of how the formalism can be applied to biophysically plausible neural network models.

Motivation & Objective

• To develop a systematic analytical framework for studying large but finite neural networks with biophysically realistic dynamics.
• To extend mean field theory by incorporating finite-size effects through higher-order statistical moments, particularly two-point correlations.
• To derive effective field equations—specifically activity equations—that describe the dynamics of moments and correlations in neural networks.
• To enable numerical simulation of these equations via lattice methods for probing non-mean-field behaviors in realistic neural systems.
• To assess the stability of neural network states under finite-N corrections, especially near critical boundaries like u(t) = 0.

Proposed method

• Formulates the neural network dynamics as a path integral using the Doi-Peliti-Janssen transformation, mapping probability density functionals into field-theoretic actions.
• Derives an effective action S = Sφ + Su, where Sφ governs phase density evolution and Su governs synaptic drive dynamics with noise and coupling.
• Applies mean field theory by minimizing the action, yielding deterministic equations for the first moments (mean phase and synaptic drive).
• Introduces activity equations via Legendre transformation, incorporating two-point correlation functions Cij to capture finite-N effects.
• Uses the 2PI (two-particle irreducible) formalism to derive closed equations for moments and correlations, including non-Markovian and noise-induced terms.
• Develops a lattice-based numerical scheme for simulating the activity equations, enabling study of correlation-driven instabilities and memory dynamics.

Experimental results

Research questions

• RQ1 How can quantum field theory techniques be adapted to model biophysically detailed neural networks with finite-size effects?
• RQ2 In what parameter regimes does mean field theory break down due to finite-N correlations in neural network dynamics?
• RQ3 How do two-point correlation functions contribute to the stability and dynamics of neural network states, particularly near critical thresholds like u(t) = 0?
• RQ4 Can the activity equations derived from the 2PI formalism provide a more accurate description of neural network behavior than mean field theory?
• RQ5 What are the implications of correlation-driven instabilities for neural memory and information processing in finite neural systems?

Key findings

• The mean field equations derived from the effective action accurately describe the first-moment dynamics of the neural network in the N → ∞ limit.
• The activity equations include a finite-N correction term proportional to 1/N, specifically involving the correlation function C31, which captures the influence of spike-induced synaptic fluctuations.
• Numerical comparisons show that the mean field solution for u(t) is stable, but the full microscopic dynamics exhibit instability near u(t) = 0, indicating breakdown of mean field predictions.
• The region where u(t) > 0 is stable under mean field analysis, while regions with u(t) < 0 rely on boundary conditions to constrain eigenvalues, suggesting sensitivity to finite-size effects.
• The activity equations reveal that correlation functions can significantly alter the dynamics, especially near phase transitions or threshold behaviors, implying that higher-order moments are essential for accurate modeling.
• The framework enables future lattice simulations to probe the role of two-point correlations in neural network stability and memory retention, particularly in non-equilibrium states.