Neural Network Field Theory (NN-FT)

• Neural Network Field Theory is a framework linking neural architectures, probabilistic ensembles, and field observables to rigorously derive free and interacting regimes.
• It maps neural dynamics into lattice models, effective actions, and continuum formulations, revealing deep connections with statistical physics and phase transitions.
• The approach provides actionable insights for optimizing training dynamics and managing finite-width corrections in modern neural network implementations.

Neural Network Field Theory (NN-FT) denotes a family of closely related frameworks in which neural networks and field theories are placed in direct correspondence. In one line of work, a field theory is defined by a neural-network architecture together with a probability density on its parameters, so that correlation functions are computed by integrating over parameter space rather than over fields. In another, random neural networks, spiking populations, or training trajectories are rewritten as lattice models, statistical field theories, or continuum actions. Across these formulations, the common structures are generating functionals, connected correlators, large-width limits, effective actions, and controlled departures from Gaussianity (Demirtas et al., 2023, Halverson, 3 Jul 2026).

1. Conceptual scope and formal definitions

A central parameter-space definition takes a network architecture

$\phi_\theta:\mathbb{R}^d\to\mathbb{R}$

together with a density $P(\theta)$ on parameters, and defines observables by

$$\langle \mathcal{O} \rangle_P = \int d\theta\, P(\theta)\,\mathcal O[\phi_\theta].$$

The corresponding $n$-point functions are

$$G_n(x_1,\dots,x_n)=\langle \phi(x_1)\cdots \phi(x_n)\rangle_P,$$

and a generating functional can be written as

$$Z[J]=\int d\theta\,P(\theta)\,\exp\!\left(\int d^dx\,J(x)\phi_\theta(x)\right).$$

In this formulation, the architecture determines the class of fields, while the parameter density determines the ensemble; if the induced Schwinger functions satisfy the Osterwalder–Schrader axioms, the construction defines a bona fide Euclidean QFT (Ferko et al., 2 Apr 2026, Halverson, 3 Jul 2026).

A second usage studies neural networks themselves with field-theoretic tools. For random feedforward networks, the distribution of preactivations can be mapped exactly onto lattice models in statistical physics, with the layer index as a one-dimensional lattice coordinate and preactivations as site variables. For stochastic rate networks and spiking networks, one instead works with generating functionals, MSRJD path integrals, effective actions, and dynamic mean-field theory to describe fluctuations, correlations, and phase structure (Schoenholz et al., 2017, 1901.10416, Qiu et al., 2014).

A third usage promotes neural-network parameters or their indices to fields. In the synaptic formulation of training dynamics, weights and biases are treated as dynamical degrees of freedom with an action principle, and in a continuum limit their indices become spatial coordinates of a field theory. This shifts the emphasis from the statistics of random networks at initialization to the geometry and dynamics of learning itself (Lee et al., 11 Mar 2025).

These usages are mathematically distinct but structurally convergent. This suggests that NN-FT is less a single formalism than a research program organized around a shared dictionary: architectures define ensembles, correlators encode observables, and continuum or large-$N$ limits expose field-theoretic structure.

2. Free theories, Gaussian limits, and lattice descriptions

For independent neurons with suitable finite-variance assumptions, the central limit theorem yields Gaussian processes in the infinite-width limit. In a single-layer network

$\phi(x)=\frac{1}{\sqrt N}\sum_{i=1}^N h_i(x),$

the connected correlators scale as

$$G_c^{(r)}(x_1,\dots,x_r)=\frac{G^{(r)}_{c,h_i}(x_1,\dots,x_r)}{N^{r/2-1}},$$

so $G_c^{(r)}\to 0$ for $P(\theta)$0 as $P(\theta)$1. The limiting theory is therefore Gaussian, with quadratic action

$P(\theta)$2

This is the basic NN-GP or generalized-free-field regime of NN-FT (Demirtas et al., 2023).

The free propagator can be engineered directly by architecture design. A canonical random Fourier feature architecture,

$P(\theta)$3

with $P(\theta)$4, $P(\theta)$5, and momentum density $P(\theta)$6, has kernel

$P(\theta)$7

Choosing

$P(\theta)$8

reproduces a standard scalar propagator $P(\theta)$9 between IR and UV cutoffs. A closely related cosine architecture for a massive scalar field is used in later NNFT analyses of finite-width bias and variance (Ferko et al., 2 Apr 2026, Zhang, 29 Apr 2026).

Random feedforward networks admit an exact statistical-mechanical rewriting in terms of preactivations. After integrating out Gaussian weights and biases, one obtains a lattice partition function over preactivation variables, with the layer index as lattice coordinate and a configuration-dependent covariance coupling adjacent layers. In the wide-network limit, both random linear networks and random rectified linear networks admit Gaussian approximations for their large-scale fluctuations; these Gaussian lattice theories can be diagonalized by discrete Fourier transform, and their long-wavelength behavior is described by an effective field theory along depth (Schoenholz et al., 2017).

At initialization, deep fully connected networks admit a complementary effective-field-theory description in which the infinite-width theory is free and finite width induces $$\langle \mathcal{O} \rangle_P = \int d\theta\, P(\theta)\,\mathcal O[\phi_\theta].$$0-suppressed interactions. The joint distribution of preactivations can be written as a Euclidean action $$\langle \mathcal{O} \rangle_P = \int d\theta\, P(\theta)\,\mathcal O[\phi_\theta].$$1, ghost fields represent determinant factors, and connected $$\langle \mathcal{O} \rangle_P = \int d\theta\, P(\theta)\,\mathcal O[\phi_\theta].$$2-point functions scale as $$\langle \mathcal{O} \rangle_P = \int d\theta\, P(\theta)\,\mathcal O[\phi_\theta].$$3. Depth acts as an RG-like direction, and a single susceptibility kernel controls the scaling of all connected correlators (Banta et al., 2023).

3. Interactions, effective actions, and collective dynamics

Beyond the Gaussian limit, NN-FT introduces interactions in two principal ways. The first is finite width: $$\langle \mathcal{O} \rangle_P = \int d\theta\, P(\theta)\,\mathcal O[\phi_\theta].$$4 corrections generate connected higher-point functions, so a free theory at $$\langle \mathcal{O} \rangle_P = \int d\theta\, P(\theta)\,\mathcal O[\phi_\theta].$$5 becomes interacting at finite $$\langle \mathcal{O} \rangle_P = \int d\theta\, P(\theta)\,\mathcal O[\phi_\theta].$$6. The second is controlled breaking of statistical independence in parameter space, which produces non-Gaussianity even at infinite width. In this setting, connected correlators can be used to reconstruct the action order by order via an Edgeworth-motivated expansion, with a diagrammatic prescription in which the vertices are connected correlators and the propagators are inverse two-point kernels. This construction is explicit enough to realize $$\langle \mathcal{O} \rangle_P = \int d\theta\, P(\theta)\,\mathcal O[\phi_\theta].$$7 theory as an infinite-$$\langle \mathcal{O} \rangle_P = \int d\theta\, P(\theta)\,\mathcal O[\phi_\theta].$$8 neural-network field theory (Demirtas et al., 2023).

A separate statistical-field-theory tradition studies interacting neural dynamics directly. For networks of rate units and binary spins, one begins from generating functionals $$\langle \mathcal{O} \rangle_P = \int d\theta\, P(\theta)\,\mathcal O[\phi_\theta].$$9, cumulant generating functions $n$0, Wick contractions for Gaussian reference theories, and MSRJD path integrals for stochastic dynamics. In random recurrent rate networks with quenched Gaussian couplings, disorder averaging and saddle-point analysis yield a dynamic mean-field theory in which each neuron is driven by a self-consistent Gaussian colored noise. The resulting autocorrelation equation recovers the Sompolinsky–Crisanti–Sommers transition between stable fixed points and chaos, and the onset of chaos is controlled by

$n$1

Loopwise effective actions and vertex functions then organize systematic corrections beyond mean field, including a diagrammatic derivation of TAP equations for pairwise maximum-entropy models (1901.10416).

Biophysical spiking networks admit yet another field-theoretic rewriting. For theta neurons with synaptic drive and noise, a Doi–Peliti / MSRJD construction yields an action $n$2 for phase-density fields $n$3, synaptic drive fields $n$4, and their response fields. Saddle-point equations reproduce population-level mean-field dynamics, while loop and 2PI corrections generate “activity equations” for means and two-point functions. In spatially structured networks with Mexican-hat coupling, mean field agrees well with microscopic simulations for small synaptic decay rate $n$5, but significant finite-$n$6 deviations appear for large $n$7, motivating correlation-corrected effective theories (Qiu et al., 2014).

This multiscale picture extends to interacting collective states. Starting from microscopic neuron and connectivity fields $n$8 and $n$9, an effective action for fluctuations $$G_n(x_1,\dots,x_n)=\langle \phi(x_1)\cdots \phi(x_n)\rangle_P,$$0 around a connectivity background can be constructed, and localized connectivity structures are then promoted to an infinite family of collective-state fields $$G_n(x_1,\dots,x_n)=\langle \phi(x_1)\cdots \phi(x_n)\rangle_P,$$1. Their interaction terms encode activation, association, and deactivation of assemblies, so the field theory describes transitions among collective neural states rather than only neuron-level dynamics (Gosselin et al., 2023).

4. Training dynamics as a synaptic field theory

In a distinct formulation, neural-network training itself is written as a field theory. Starting from gradient descent for weights and biases,

$$G_n(x_1,\dots,x_n)=\langle \phi(x_1)\cdots \phi(x_n)\rangle_P,$$2

one embeds the first-order flow into a second-order dissipative equation

$$G_n(x_1,\dots,x_n)=\langle \phi(x_1)\cdots \phi(x_n)\rangle_P,$$3

which follows from the action

$$G_n(x_1,\dots,x_n)=\langle \phi(x_1)\cdots \phi(x_n)\rangle_P,$$4

In the large-friction regime $$G_n(x_1,\dots,x_n)=\langle \phi(x_1)\cdots \phi(x_n)\rangle_P,$$5, this reduces order by order to ordinary gradient descent (Lee et al., 11 Mar 2025).

The field-theoretic step is a continuum limit in the parameter indices. Instead of discrete parameters $$G_n(x_1,\dots,x_n)=\langle \phi(x_1)\cdots \phi(x_n)\rangle_P,$$6 and $$G_n(x_1,\dots,x_n)=\langle \phi(x_1)\cdots \phi(x_n)\rangle_P,$$7, one introduces a field $$G_n(x_1,\dots,x_n)=\langle \phi(x_1)\cdots \phi(x_n)\rangle_P,$$8 or $$G_n(x_1,\dots,x_n)=\langle \phi(x_1)\cdots \phi(x_n)\rangle_P,$$9, where the layer, neuron, or connectivity indices are reinterpreted as coordinates $$Z[J]=\int d\theta\,P(\theta)\,\exp\!\left(\int d^dx\,J(x)\phi_\theta(x)\right).$$0 in an emergent space. In a toy model with one-dimensional local connectivity and linear activation, the discrete cost can be reorganized as

$$Z[J]=\int d\theta\,P(\theta)\,\exp\!\left(\int d^dx\,J(x)\phi_\theta(x)\right).$$1

with $$Z[J]=\int d\theta\,P(\theta)\,\exp\!\left(\int d^dx\,J(x)\phi_\theta(x)\right).$$2, $$Z[J]=\int d\theta\,P(\theta)\,\exp\!\left(\int d^dx\,J(x)\phi_\theta(x)\right).$$3, and $$Z[J]=\int d\theta\,P(\theta)\,\exp\!\left(\int d^dx\,J(x)\phi_\theta(x)\right).$$4 determined by the training data. In the continuum this becomes

$$Z[J]=\int d\theta\,P(\theta)\,\exp\!\left(\int d^dx\,J(x)\phi_\theta(x)\right).$$5

leading to the action

$$Z[J]=\int d\theta\,P(\theta)\,\exp\!\left(\int d^dx\,J(x)\phi_\theta(x)\right).$$6

Training data therefore appear as external sources in the form of position-dependent couplings $$Z[J]=\int d\theta\,P(\theta)\,\exp\!\left(\int d^dx\,J(x)\phi_\theta(x)\right).$$7 and $$Z[J]=\int d\theta\,P(\theta)\,\exp\!\left(\int d^dx\,J(x)\phi_\theta(x)\right).$$8 (Lee et al., 11 Mar 2025).

The time-dependent measure factor has a geometric interpretation. Comparing

$$Z[J]=\int d\theta\,P(\theta)\,\exp\!\left(\int d^dx\,J(x)\phi_\theta(x)\right).$$9

with $N$0 identifies the effective spacetime as de Sitter in flat slicing, with

$N$1

Training time $N$2 becomes cosmic time, and the step size $N$3 controls the Hubble parameter. In this sense, standard gradient descent is the overdamped limit of a synaptic field evolving in an expanding spacetime (Lee et al., 11 Mar 2025).

The construction is intentionally limited. The explicit continuum example uses linear activation, a single layer, and local connectivity; the analysis does not compute functional integrals over weights, perform RG analyses, or connect quantitatively to generalization error. The authors emphasize that for realistic deep architectures the induced synaptic field theory is expected to be strongly nonlocal and nonpolynomial.

5. Symmetry, topology, conformal structure, and strings

NN-FT can incorporate topological sectors by enlarging latent space with discrete variables. For mixed continuous/discrete parameters $N$4,

$N$5

where $N$6 may represent winding, momentum–winding pairs, or vortex configurations. In the two-dimensional compact boson, the field is decomposed as

$N$7

with a random-Fourier-feature spin-wave sector

$N$8

and an explicit vortex sector

$N$9

The spin-wave correlator gives

$\phi(x)=\frac{1}{\sqrt N}\sum_{i=1}^N h_i(x),$0

so the Gaussian fixed line has critical point

$\phi(x)=\frac{1}{\sqrt N}\sum_{i=1}^N h_i(x),$1

Including vortices reproduces the full BKT phenomenology: algebraic decay below $\phi(x)=\frac{1}{\sqrt N}\sum_{i=1}^N h_i(x),$2, exponential decay above $\phi(x)=\frac{1}{\sqrt N}\sum_{i=1}^N h_i(x),$3, the essential singularity

$\phi(x)=\frac{1}{\sqrt N}\sum_{i=1}^N h_i(x),$4

and the universal helicity-modulus jump (Ferko et al., 2 Apr 2026).

The same mixed-ensemble logic realizes T-duality for compact bosons. For a circle of radius $\phi(x)=\frac{1}{\sqrt N}\sum_{i=1}^N h_i(x),$5, one samples a non-compact oscillator network together with discrete momentum and winding labels $\phi(x)=\frac{1}{\sqrt N}\sum_{i=1}^N h_i(x),$6, with the duality

$\phi(x)=\frac{1}{\sqrt N}\sum_{i=1}^N h_i(x),$7

The framework reproduces not only circle duality but also Buscher rules on constant toroidal backgrounds, enhanced $\phi(x)=\frac{1}{\sqrt N}\sum_{i=1}^N h_i(x),$8 current algebra at the self-dual radius, and non-geometric T-fold transition functions (Ferko et al., 2 Apr 2026).

A separate development shows that NN-FT can realize the full Virasoro symmetry of a two-dimensional CFT. Generic conformal NN-FT constructions based on generalized free fields lack a local stress tensor satisfying conformal Ward identities, so they do not realize local conformal symmetry. This obstruction is overcome by a log-kernel architecture whose infinite-width limit is a neural free boson with

$\phi(x)=\frac{1}{\sqrt N}\sum_{i=1}^N h_i(x),$9

holomorphic current $$G_c^{(r)}(x_1,\dots,x_r)=\frac{G^{(r)}_{c,h_i}(x_1,\dots,x_r)}{N^{r/2-1}},$$0, and local stress tensor

$$G_c^{(r)}(x_1,\dots,x_r)=\frac{G^{(r)}_{c,h_i}(x_1,\dots,x_r)}{N^{r/2-1}},$$1

Numerically, the stress-tensor correlator yields

$$G_c^{(r)}(x_1,\dots,x_r)=\frac{G^{(r)}_{c,h_i}(x_1,\dots,x_r)}{N^{r/2-1}},$$2

consistent with $$G_c^{(r)}(x_1,\dots,x_r)=\frac{G^{(r)}_{c,h_i}(x_1,\dots,x_r)}{N^{r/2-1}},$$3, while neural vertex operators $$G_c^{(r)}(x_1,\dots,x_r)=\frac{G^{(r)}_{c,h_i}(x_1,\dots,x_r)}{N^{r/2-1}},$$4 reproduce the free-boson scaling dimensions $$G_c^{(r)}(x_1,\dots,x_r)=\frac{G^{(r)}_{c,h_i}(x_1,\dots,x_r)}{N^{r/2-1}},$$5. The same framework extends to a neural Majorana fermion with $$G_c^{(r)}(x_1,\dots,x_r)=\frac{G^{(r)}_{c,h_i}(x_1,\dots,x_r)}{N^{r/2-1}},$$6, an $$G_c^{(r)}(x_1,\dots,x_r)=\frac{G^{(r)}_{c,h_i}(x_1,\dots,x_r)}{N^{r/2-1}},$$7 scalar multiplet with $$G_c^{(r)}(x_1,\dots,x_r)=\frac{G^{(r)}_{c,h_i}(x_1,\dots,x_r)}{N^{r/2-1}},$$8, and conformal or superconformal boundary conditions via the method of images (Robinson, 30 Dec 2025).

The parameter-space formalism also supports a generalized Ward-identity and anomaly calculus. A master identity derived from total derivatives in parameter space yields Schwinger–Dyson equations and a breaking function

$$G_c^{(r)}(x_1,\dots,x_r)=\frac{G^{(r)}_{c,h_i}(x_1,\dots,x_r)}{N^{r/2-1}},$$9

with explicit-breaking and Jacobian components. Applied to the neural realization of worldsheet bosons and ghosts on the sphere, the regularized Weyl anomaly is

$G_c^{(r)}\to 0$0

and demanding vanishing anomaly recovers the bosonic-string critical dimension $G_c^{(r)}\to 0$1. The same NN-FT architecture reproduces the Koba–Nielsen factor and the Virasoro–Shapiro amplitude by integrating directly over network parameters (Halverson, 3 Jul 2026).

6. Finite-width accuracy, architectural optimization, and open problems

Although many NN-FT constructions are exact at infinite width, finite-$G_c^{(r)}\to 0$2 effects are often decisive in practice. For a massive scalar field represented by a one-hidden-layer cosine network,

$G_c^{(r)}\to 0$3

there is a one-parameter family of architectures indexed by $G_c^{(r)}\to 0$4 that all reproduce the same free propagator as $G_c^{(r)}\to 0$5. The splitting

$G_c^{(r)}\to 0$6

leaves the infinite-width theory invariant but changes finite-width bias and variance. For the massive scalar, $G_c^{(r)}\to 0$7 corresponds to propagator-weighted neuron momenta and constant neuron amplitudes, minimizes finite-width variance, and uniquely removes IR-sensitive corrections in the interacting theory. Even then, relative errors from both bias and variance grow exponentially with distance beyond the correlation length, bias can be removed by extrapolating to infinite width, and the variance imposes a signal-to-noise bound analogous to that of lattice field theory (Zhang, 29 Apr 2026).

A related perturbative analysis of scalar $G_c^{(r)}\to 0$8 theory implemented on neurons reaches a similar conclusion from the QFT side. Breaking statistical independence of network parameters can realize local interactions exactly at infinite neuron number, but the renormalized $G_c^{(r)}\to 0$9 corrections to two- and four-point correlators are sensitive to the ultraviolet cutoff and therefore have weak convergence. A modified architecture that keeps only interaction terms with all four fields at a vertex coming from distinct neurons removes one-loop bubble and non-Gaussian contributions at that order and improves the scaling of perturbative corrections, but higher-order non-Gaussian effects remain (Sen et al., 5 Aug 2025).

Initialization-side EFTs identify another finite-width issue: criticality is not a property of the two-point kernel alone. In the diagrammatic effective theory of deep fully connected networks, the same susceptibility kernel governs the RG flow of every connected correlator, and the criticality condition

$P(\theta)$00

implies that all connected correlators display power-law rather than exponential scaling with depth (Banta et al., 2023).

Several recurring misconceptions are therefore addressed by the literature itself. Infinite width does not by itself guarantee locality, a local stress tensor, or a full conformal algebra; generic conformal NN-FTs remain generalized free fields (Robinson, 30 Dec 2025). Topological sectors do not emerge automatically from smooth Gaussian architectures; they are introduced explicitly through discrete latent variables (Ferko et al., 2 Apr 2026). Training-dynamics field theories provide an action principle and a geometric interpretation of optimization, but they do not yet furnish quantitative accounts of empirical generalization in realistic deep networks (Lee et al., 11 Mar 2025).

Open problems are correspondingly concrete. The literature repeatedly points to continuum synaptic field theories for realistic architectures such as FCNs, CNNs, ResNets, and transformers; nonlocal interactions in weight space; extensions of NN-FT to gauge theories, fermions, higher-form symmetries, and defects; systematic variance-reduction methods for parameter-space QFT simulation; interacting two-dimensional CFTs beyond free bosons and Majorana fermions; and trained rather than purely initialized field theories (Lee et al., 11 Mar 2025, Zhang, 29 Apr 2026, Robinson, 30 Dec 2025, Halverson, 3 Jul 2026).

In aggregate, NN-FT is best understood as a set of mathematically explicit correspondences rather than a single canonical model. Its unifying claim is that neural architectures, probability measures on parameters, and field-theoretic observables can be organized into a common formal language. Within that language, free theories arise from Gaussian-process limits, interactions from finite width or correlated priors, topological sectors from discrete latent variables, symmetries from architectural and measure-theoretic constraints, and training itself from an action on synaptic degrees of freedom.