Pith. sign in

REVIEW 2 major objections 5 minor 177 references

A field theory can be defined by a neural-network architecture and a density on its parameters, recovering free fields, Liouville, string amplitudes, BKT, and D=26.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-12 06:13 UTC pith:X5WCTNN4

load-bearing objection Solid lecture notes that package the 2025–26 NN-FT papers into a usable three-day course; the value is synthesis and pedagogy, not a new theorem. the 2 major comments →

arxiv 2607.02905 v1 pith:X5WCTNN4 submitted 2026-07-03 hep-th

Pre-Strings Lectures on Artificial Intelligence

James Halverson This is my paper
classification hep-th
keywords neural network field theoryNNGPLiouville theorybosonic stringcritical dimensionBKT transitionWard identitiesCalabi-Yau metrics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

These lecture notes argue that the data of a neural network—its architecture together with a probability density on its parameters—already define a field theory, because correlators can be computed by integrating over parameters rather than over field configurations. Infinite-width networks become free fields by the central limit theorem; interactions appear when that theorem is broken by finite width or by deforming the density. With this definition the notes reconstruct Liouville structure constants, the Virasoro–Shapiro and Veneziano amplitudes, the BKT transition via discrete topological latents, Ward identities and anomalies on parameter space, and a mode-counting derivation of the bosonic-string critical dimension D=26. They also survey how the same tools reverse: physics-informed networks for Calabi–Yau metrics, reinforcement learning over string vacua and knots, and interpretable models aimed at conjecture generation. A sympathetic reader cares because the construction supplies both a new language for field theory and concrete computational routes to quantities that usually require path integrals or lattice methods.

Core claim

The central claim is that a neural-network field theory is the pair (architecture, parameter density), with the partition function an integral over parameters. From this definition one engineers free theories by the CLT and spectrum shaping, interactions by 1/N corrections or independence-breaking deformations, and quantum theories by clearing Osterwalder–Schrader conditions. Concrete ensembles then recover Liouville DOZZ constants, string amplitudes, BKT phenomenology, Ward identities, and a Jacobian mode-count that yields D=26.

What carries the argument

Neural Network Field Theory (NN-FT): the data (ϕ_θ, P(θ)) whose parameter-space partition function Z[J] = ∫ dθ P(θ) exp(∫ J ϕ_θ) defines all correlators. Free limits follow from the CLT; interactions from finite N or density deformations; symmetries and anomalies from flows on parameter space that leave or break that integral.

Load-bearing premise

That the Euclidean correlators of these engineered ensembles continue to unitary Lorentzian quantum field theories, which requires reflection positivity and cluster decomposition that are verified only in free and quantum-mechanics cases so far.

What would settle it

Construct an explicit finite-width NN-FT whose four-point and higher correlators can be computed exactly, check whether the Osterwalder–Schrader reflection-positivity inequalities hold after analytic continuation, and compare the resulting Lorentzian spectrum or S-matrix against a known continuum QFT or string amplitude.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Any Euclidean QFT in the constructive sense admits an NN-FT description with countably many parameters (universality).
  • Liouville three-point structure constants can be obtained by Monte Carlo sampling of a conditioned zero-mode density plus free spherical-harmonic modes.
  • Virasoro–Shapiro and Veneziano amplitudes arise as exact finite-dimensional integrals over network parameters once free bosons and ghosts are realized as infinite-width ensembles.
  • Topological sectors enter as discrete latent variables, allowing vortex unbinding and the BKT jump to be read off from network correlators.
  • The critical dimension D=26 follows from a regularized mismatch between bosonic and Grassmann Jacobians under a Weyl flow on parameter space.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If reflection positivity can be engineered systematically, NN-FT becomes a constructive route to interacting QFTs that bypasses the need to write an action first.
  • The same parameter-space Ward identity machinery that produced D=26 should apply to other anomaly coefficients (central charges, chiral anomalies) once the relevant flows are identified.
  • Agentic verification loops wrapped around NN-FT correlator codes could turn the Monte-Carlo Liouville and BKT checks into automated, continuously refined numerical proofs.
  • Architecture-robust free-boson constructions suggest that finite-N non-Gaussian corrections to string amplitudes are themselves architecture-dependent observables worth classifying.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. These lecture notes present a three-day course on AI for string theorists. Day 1 develops neural-network essentials (expressivity, statistics via NNGP and 1/N non-Gaussianities, dynamics via NTK and µP) with a field-theoretic vocabulary. Day 2 defines Neural Network Field Theory (NN-FT) as the data of an architecture ϕ_θ together with a density P(θ), with correlators computed from the parameter-space partition function; free theories arise by CLT plus spectrum shaping, interactions by CLT violation, and recent constructions recover Liouville DOZZ constants, Virasoro-Shapiro/Veneziano amplitudes, BKT phenomenology with vortices, Ward identities/anomalies, and a mode-counting derivation of D=26. Day 3 surveys applied ML for strings (agents, PINNs for Calabi-Yau metrics, RL on vacua and unknotting, conjecture generation). The notes synthesize and pedagogically reorganize results largely already posted as research papers.

Significance. If the NN-FT definition and constructions hold as stated, the notes supply a coherent, field-theoretic language for ensembles of networks and a concrete engineering toolkit (CLT free theories, independence-breaking interactions, absorption for symmetries, discrete latents for topology) that recovers standard results of 2d CFT, string theory, and topological phase transitions. Strengths include analytic recovery of string amplitudes from Gaussian parameter integrals, percent-level numerical DOZZ agreement with reported error bars, an explicit zeta-regularized Jacobian count yielding D=26, and open acknowledgment that OS axioms beyond free/QM cases remain open. As lecture notes they are valuable for training and for consolidating a rapidly growing literature; they do not claim a single new theorem beyond the cited papers.

major comments (2)
  1. §3.1.3 (and the OS checklist): the claim that NN ensembles define quantum field theories rests on Osterwalder–Schrader reconstruction. The notes correctly state that reflection positivity and cluster decomposition are verified only for free cases and quantum mechanics, and remain open for the interacting constructions (Liouville, strings, BKT). This is load-bearing for any Lorentzian/unitary reading of Day 2. The Euclidean constructions and amplitude recoveries stand independently, but the manuscript should either (i) restrict the “QFT” language more carefully to Euclidean correlators throughout §§3.2.2–3.2.5, or (ii) supply at least a sketch of how RP/cluster are expected to hold (or fail) for the deformed densities used in Liouville and ϕ^4.
  2. §3.2.3, Eqs. (156)–(165) and the D=26 count in §3.2.5, Eqs. (189)–(194): the free-boson and ghost architectures are engineered so that the two-point functions match the standard worldsheet propagators (spectrum shaping + Grassmann Gaussians). Once that is fixed, Virasoro-Shapiro/Veneziano and the critical-dimension count follow by standard Gaussian integrals and zeta regularization of mode Jacobians. This is design, not circular derivation of the target numbers, but the notes should state more explicitly that the physical content is the choice of architecture/density that realizes the free theory, after which the amplitudes and anomaly are recovered rather than independently predicted. A short paragraph distinguishing “engineering the free theory” from “deriving the amplitude” would prevent misreading.
minor comments (5)
  1. Introduction and §5: the notes are dated July 2026 and cite contemporaneous results (IMO gold, Erdős counterexamples, agentic systems). For archival publication, a brief note on the snapshot nature of the “current AI landscape” paragraphs would help future readers.
  2. Fig. 3 (Liouville DOZZ): the caption reports L=30, 10×50k runs, and error bars smaller than markers. Adding a short statement of the Monte-Carlo estimator variance or a reference to the companion paper’s numerical appendix would strengthen reproducibility claims.
  3. §2.4.3 (µP scaling): the one-parameter family and the unique solution with η∼O_N(1) are dense. A small table of (a_ℓ,b_ℓ,c,d) for the standard NTK vs µP regimes would aid readers who skip the index algebra.
  4. §4.1 (agents): the GPD pipeline and “order-of-magnitude drop in implementation barrier” are useful but anecdotal. A single concrete before/after example (e.g., time to reproduce a Day-2 baseline) would make the claim more falsifiable.
  5. Typos and polish: the disclaimer already notes residual typos from rapid posting; a pass for consistency of notation (e.g., θ vs θ(t), G^{(n)} vs G_c^{(4)}) and for missing cross-references between Day-1 NTK and Day-3 metric flows would improve readability.

Circularity Check

0 steps flagged

No load-bearing circularity: free theories are engineered by design (CLT + spectrum shaping), after which amplitudes, DOZZ match, BKT exponents, and the D=26 mode count are independent analytic or numerical checks against external benchmarks.

full rationale

The manuscript is lecture notes whose core definition (NN-FT = architecture ϕ_θ plus density P(θ), Def. 3.1) is stipulative, not a derivation of a prior claim. Free theories are obtained by the CLT plus explicit spectrum-shaping of G^{(2)} (Scalar-net, free-boson random features, ghosts); this is engineering, openly stated as a recipe, not a prediction of the free propagator. Once the free ensemble is fixed, the Liouville Monte-Carlo match to DOZZ, the analytic recovery of Virasoro-Shapiro/Veneziano from parameter-space Gaussian integrals, the BKT power-law/essential-singularity/stiffness-jump phenomenology, and the zeta-regularized Jacobian count that yields D=26 are all independent of the free-theory input and are checked against external formulas or known results. The universality theorem invokes the Borel isomorphism theorem (external mathematics) for existence of a (possibly non-constructive) architecture. Self-citations to the author's recent papers supply details of the constructions but are not used as unverified uniqueness theorems that force the target numbers; the sketches given in the notes are self-contained enough to exhibit the reductions. The only openly acknowledged gap (OS reflection positivity beyond free/QM cases) is orthogonal to circularity. Score 1 only for the ordinary presence of author-overlapping citations that are not load-bearing for the claimed recoveries.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

NN-FT rests on the definition that a field theory is an architecture plus a parameter density, the CLT for free limits, the Osterwalder–Schrader reconstruction for quantumness, and standard QFT/ML background (Ward identities via change of variables, zeta regularization, calibrated geometry for CY volumes). Free parameters appear as architectural scales (width N, weight variances, cutoffs) that are fixed to match known free propagators or physical constants such as α′. No new particles or forces are postulated; the main invented entity is the NN-FT definition itself.

free parameters (4)
  • output-weight variance σ_a (or equivalent)
    Sets the string tension α′ via α′ = σ_a^{2}/(Λ^{2}−ϵ^{2}) in the free-boson architecture; chosen to match the desired physical value.
  • width N and UV/IR cutoffs Λ, ϵ
    Control the continuum and infinite-width limits; regulators that must be sent to infinity after correlators are computed.
  • Liouville zero-mode conditioning parameters (b, μ, α_i)
    Physical Liouville parameters; sampled or fixed to compare against DOZZ, not fitted to produce the match.
  • vortex fugacity y and stiffness K_0 (or b)
    Control the BKT temperature; scanned to locate the transition rather than fitted post hoc to force η=1/4.
axioms (5)
  • standard math Central Limit Theorem in function space: sum of N i.i.d. neurons becomes a Gaussian process as N→∞
    Used throughout Day 1–2 to obtain free NN-FTs (NNGP correspondence).
  • domain assumption Osterwalder–Schrader reconstruction: Euclidean correlators obeying the OS axioms define a unitary Lorentzian QFT
    Invoked in §3.1.3 to claim that certain NN ensembles are quantum; reflection positivity remains open beyond QM.
  • standard math Borel isomorphism theorem: any two standard Borel spaces are isomorphic, so every probability measure on S′(R^d) is the push-forward of a parameter measure
    Underpins the universality theorem (Theorem 3.3).
  • domain assumption Yau’s theorem: a compact Kähler manifold with c1=0 admits a unique Ricci-flat metric in each Kähler class
    Background for the PINN/Calabi–Yau section; existence is assumed, the network approximates the metric.
  • ad hoc to paper Absorption mechanism: a symmetry of the ensemble exists when a field transformation can be absorbed into a reparameterization that leaves dθ P(θ) invariant
    Core technical device of Day 1–2 for engineering global symmetries and later Ward identities; introduced in the author’s prior work and used as definitional here.
invented entities (1)
  • Neural Network Field Theory (NN-FT) independent evidence
    purpose: Defines a field theory by the pair (architecture ϕ_θ, density P(θ)) with partition function an integral over parameters rather than fields.
    The central conceptual object of Day 2; independent evidence is the recovery of known correlators and amplitudes, but the definition itself is a modeling choice.

pith-pipeline@v1.1.0-grok45 · 51799 in / 3204 out tokens · 36874 ms · 2026-07-12T06:13:12.695379+00:00 · methodology

0 comments
read the original abstract

These notes are based on six lectures given over three days at the Pre-Strings 2026 school in Shanghai. Day 1 develops neural network essentials, organized around the expressivity, statistics, and dynamics of neural networks, presented with a field-theoretic lens. Day 2 develops a neural network approach to field theory (NN-FT), in which a field theory is defined by a network architecture and a density on its parameters, and surveys recent results. Examples include a universality theorem, a neural network realization of Liouville theory, famous string amplitudes, topological sectors and the Kosterlitz-Thouless transition, Ward identities and anomalies, and a new derivation of the critical dimension of the bosonic string. Day 3 turns the lens around and covers applied AI for string theory: agentic workflows that are changing how the other techniques are implemented, physics-informed neural networks and Calabi-Yau metrics, reinforcement learning and search in the string landscape and in knot theory, and interpretable supervised learning with an eye towards conjecture generation.

Figures

Figures reproduced from arXiv: 2607.02905 by James Halverson.

Figure 1
Figure 1. Figure 1: The mechanism behind the Universal Approximation The [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: MLPs versus KANs: functional forms, the mathematical [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The Liouville structure constant C(α1, α2, α3) from NN correlators (points; α1 = α2 fixed and α3 varied, for several b) versus the exact DOZZ formula (solid curves), which lies within all error bars. NN-FT run: L = 30 spherical harmonics, 10 experiments of 50,000 runs, sphere pixelated at 100 × 200 points; the error bars (variance across experiments) are smaller than the markers [67]. the free-boson logari… view at source ↗
Figure 4
Figure 4. Figure 4: Vortex density ρv versus the coupling b (the dashed line marks bc = 1 2 ): consistent with zero below the transition, rising steeply above it as vortex–antivortex pairs unbind. From [69]. the integral and its measure are invariant, and read off a conservation law when they are, or an anomaly when the measure is not. The result of [60] is that NN-FT has its own integral, over pa￾rameters, and the same chang… view at source ↗
Figure 5
Figure 5. Figure 5: The agent loop. A reasoning model repeatedly thinks, acts [PITH_FULL_IMAGE:figures/full_fig_p030_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Hierarchy of neural-network metric flows [152]. The gen [PITH_FULL_IMAGE:figures/full_fig_p035_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The Artin generator σ1 ∈ B2: two strands with one cross￾ing. Closing the braid (joining each strand’s top to its own bottom) gives a single loop, the unknot. where destabilization, free cancellation, and the braid relation are the braid-word avatars of Reidemeister moves I, II, and III. An unknot is certified by a sequence of these reducing the word to the empty one. Cast move-selection as an MDP and train… view at source ↗
Figure 8
Figure 8. Figure 8: The braid relation σiσi+1σi = σi+1σiσi+1 (here i=1 on three strands), the Reidemeister-III move in braid-word form, and the move the trained agent relies on most. The remaining move types act on the word the same algebraic way: far commutation σiσj = σjσi , free cancellation σiσ −1 i = 1 (Reidemeister II), and the Markov (de)stabilization (Reidemeister I). 4.4 Conjecture Generation and Rigor 4.4.1 From Pre… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

177 extracted references · 13 canonical work pages · 9 internal anchors

  1. [1]

    Halverson,TASI lectures on physics for machine learning,

    J. Halverson,TASI lectures on physics for machine learning,

  2. [2]

    https://arxiv.org/abs/2408.00082

    Pith/arXiv arXiv

  3. [3]

    Isaacson,The Innovators: How a Group of Hackers, Geniuses, and Geeks Created the Digital Revolution

    W. Isaacson,The Innovators: How a Group of Hackers, Geniuses, and Geeks Created the Digital Revolution. Simon & Schuster, New York, first simon & schuster hardcover edition ed., 2014

    2014

  4. [4]

    Jaech, A

    OpenAI, :, A. Jaech, A. Kalai, A. Lerer, A. Richardson et al., OpenAI o1 system card, 2026. https://arxiv.org/abs/2412.16720

    Pith/arXiv arXiv 2026

  5. [5]

    DeepSeek-AI, D. Guo, D. Yang, H. Zhang, J. Song, P. Wang et al.,DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning, 2026. https://doi.org/10.1038/s41586-025-09422-z, https://arxiv.org/abs/2501.12948

    Pith/arXiv arXiv doi:10.1038/s41586-025-09422-z 2026

  6. [6]

    https://deepmind.google/discover/blog/advanced-version-of- gemini-with-deep-think-officially-achieves-gold-medal- standard-at-the-international-mathematical-olympiad/

    Google DeepMind,Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the international mathematical olympiad, July, 2025. https://deepmind.google/discover/blog/advanced-version-of- gemini-with-deep-think-officially-achieves-gold-medal- standard-at-the-international-mathematical-olympiad/

    2025

  7. [7]

    OpenAI model earns gold-medal score at international math olympiad and advances path to artificial general intelligence

    D. E. B´ echard, “OpenAI model earns gold-medal score at international math olympiad and advances path to artificial general intelligence.” Scientific American, 2025. https://www.scientificamerican.com/article/openai-model- earns-gold-medal-score-at-international-math-olympiad-and/

    2025

  8. [8]

    The Erd˝ os problems project

    T. F. Bloom, “The Erd˝ os problems project.” https://www.erdosproblems.com

  9. [9]

    https://openai.com/index/model- disproves-discrete-geometry-conjecture/

    OpenAI,An OpenAI model has disproved a central conjecture in discrete geometry, 2026. https://openai.com/index/model- disproves-discrete-geometry-conjecture/. Pre-Strings Lectures on Artificial Intelligence 43

    2026

  10. [10]

    N. Alon, T. F. Bloom, W. T. Gowers, D. Litt, W. Sawin, A. Shankar et al.,Remarks on the disproof of the unit distance conjecture,2605.20695

    Pith/arXiv arXiv

  11. [11]

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan et al.,React: Synergizing reasoning and acting in language models, 2023. https://arxiv.org/abs/2210.03629

    Pith/arXiv arXiv 2023

  12. [12]

    https://claude.com/claude-code

    Anthropic,Claude Code, 2025. https://claude.com/claude-code

    2025

  13. [13]

    Romera-Paredes, M

    B. Romera-Paredes, M. Barekatain, A. Novikov, M. Balog, M. P. Kumar, E. Dupont et al.,Mathematical discoveries from program search with large language models,Nature625 (2024) 468

    2024

  14. [14]

    Novikov, N

    A. Novikov, N. V˜ u, M. Eisenberger, E. Dupont, P.-S. Huang, A. Z. Wagner et al.,Alphaevolve: A coding agent for scientific and algorithmic discovery, 2025. https://arxiv.org/abs/2506.13131. [14]Pre-strings school 2026, 2026. https://www.pre-strings2026.com/

    Pith/arXiv arXiv 2025

  15. [15]

    A Stochastic Approximation Method

    H. Robbins and S. Monro, “A Stochastic Approximation Method.” Annals of Mathematical Statistics, 1951. 10.1214/aoms/1177729586, https://doi.org/10.1214/aoms/1177729586

    doi:10.1214/aoms/1177729586 1951

  16. [16]

    Rosenblatt,The perceptron: a probabilistic model for information storage and organization in the brain., Psychological review65 6(1958) 386

    F. Rosenblatt,The perceptron: a probabilistic model for information storage and organization in the brain., Psychological review65 6(1958) 386

    1958

  17. [17]

    D. P. Kingma and J. Ba,Adam: A method for stochastic optimization, 2017. https://arxiv.org/abs/1412.6980

    Pith/arXiv arXiv 2017

  18. [18]

    G. B. D. Luca and E. Silverstein,Born-infeld (BI) for AI: Energy-conserving descent (ECD) for optimization, 2022. https://arxiv.org/abs/2201.11137

    Pith/arXiv arXiv 2022

  19. [19]

    G. B. D. Luca, A. Gatti and E. Silverstein,Improving energy conserving descent for machine learning: Theory and practice,

  20. [20]

    https://arxiv.org/abs/2306.00352

    Pith/arXiv arXiv

  21. [21]

    Approximation by Superpositions of a Sigmoidal Function

    G. Cybenko, “Approximation by Superpositions of a Sigmoidal Function.” Mathematics of Control, Signals and Systems, 1989. 10.1007/BF02551274, https://doi.org/10.1007/bf02551274

    doi:10.1007/bf02551274 1989

  22. [22]

    Hornik, M

    K. Hornik, M. Stinchcombe and H. White,Approximation capabilities of multilayer feedforward networks,Neural Networks4(1991) 251

    1991

  23. [23]

    A. N. Kolmogorov,On the representation of continuous functions of several variables by superpositions of continuous functions of one variable and addition,Doklady Akademii Nauk SSSR114(1957) 953

    1957

  24. [24]

    V. I. Arnold,On functions of three variables,Doklady Akademii Nauk SSSR114(1957) 679

    1957

  25. [25]

    Z. Liu, Y. Wang, S. Vaidya, F. Ruehle, J. Halverson, M. Soljaˇ ci´ c et al.,Kan: Kolmogorov-arnold networks, 2025. https://arxiv.org/abs/2404.19756

    Pith/arXiv arXiv 2025

  26. [26]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez et al.,Attention is all you need, 2023. https://arxiv.org/abs/1706.03762

    Pith/arXiv arXiv 2023

  27. [27]

    T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal et al.,Language models are few-shot learners,

  28. [28]

    https://arxiv.org/abs/2005.14165

    Pith/arXiv arXiv 2005

  29. [29]

    C. Yun, S. Bhojanapalli, A. S. Rawat, S. J. Reddi and S. Kumar,Are transformers universal approximators of sequence-to-sequence functions?, 2020. https://arxiv.org/abs/1912.10077. Pre-Strings Lectures on Artificial Intelligence 44

    Pith/arXiv arXiv 2020

  30. [30]

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia et al.,Chain-of-thought prompting elicits reasoning in large language models, 2023. https://arxiv.org/abs/2201.11903

    Pith/arXiv arXiv 2023

  31. [31]

    Merrill and A

    W. Merrill and A. Sabharwal,The expressive power of transformers with chain of thought, 2024. https://arxiv.org/abs/2310.07923

    Pith/arXiv arXiv 2024

  32. [32]

    G. Feng, B. Zhang, Y. Gu, H. Ye, D. He and L. Wang, Towards revealing the mystery behind chain of thought: A theoretical perspective, 2023. https://arxiv.org/abs/2305.15408

    Pith/arXiv arXiv 2023

  33. [33]

    R. M. Neal,Bayesian learning for neural networks,Lecture Notes in Statistics118(1996)

    1996

  34. [34]

    C. K. Williams,Computing with infinite networks,Advances in neural information processing systems(1996)

    1996

  35. [35]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever and G. E. Hinton,Imagenet classification with deep convolutional neural networks, in Advances in Neural Information Processing Systems (F. Pereira, C. Burges, L. Bottou and K. Weinberger, eds.), vol. 25, Curran Associates, Inc., 2012, https://proceedings.neurips.cc/paper - files/paper/2012/file/c399862d3b9d6b76c8436e924a68c45b-...

    2012

  36. [36]

    Yang,Tensor programs i: Wide feedforward or recurrent neural networks of any architecture are gaussian processes,

    G. Yang,Tensor programs i: Wide feedforward or recurrent neural networks of any architecture are gaussian processes,

  37. [37]

    https://arxiv.org/abs/1910.12478

    Pith/arXiv arXiv 1910

  38. [38]

    Demirtas, J

    M. Demirtas, J. Halverson, A. Maiti, M. D. Schwartz and K. Stoner,Neural network field theories: Non-gaussianity, actions, and locality, 2023. https://arxiv.org/abs/2307.03223

    Pith/arXiv arXiv 2023

  39. [39]

    T. S. Cohen and M. Welling,Group equivariant convolutional networks, 2016. https://arxiv.org/abs/1602.07576

    Pith/arXiv arXiv 2016

  40. [40]

    Winkels and T

    M. Winkels and T. S. Cohen,3D G-CNNs for pulmonary nodule detection, 2018. https://arxiv.org/abs/1804.04656

    Pith/arXiv arXiv 2018

  41. [41]

    Kaplan, S

    J. Kaplan, S. McCandlish, T. Henighan, T. B. Brown, B. Chess, R. Child et al.,Scaling laws for neural language models, 2020. https://arxiv.org/abs/2001.08361

    Pith/arXiv arXiv 2020

  42. [42]

    E(3)-Equivariant Graph Neural Networks for Data-Efficient and Accurate Interatomic Potentials

    S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth et al.,E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials, 2021. https://doi.org/10.1038/s41467-022-29939-5, https://arxiv.org/abs/2101.03164

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1038/s41467-022-29939-5 2021

  43. [43]

    N. Frey, R. Soklaski, S. Axelrod, S. Samsi, R. Gomez-Bombarelli, C. Coley et al.,Neural scaling of deep chemical models,ChemRxiv(2022)

    2022

  44. [44]

    Boyda, G

    D. Boyda, G. Kanwar, S. Racani` ere, D. J. Rezende, M. S. Albergo, K. Cranmer et al.,Sampling usingsu(n)gauge equivariant flows, 2020. https://doi.org/10.1103/PhysRevD.103.074504, https://arxiv.org/abs/2008.05456

    Pith/arXiv arXiv doi:10.1103/physrevd.103.074504 2020

  45. [45]

    Maiti, K

    A. Maiti, K. Stoner and J. Halverson,Symmetry-via-duality: Invariant neural network densities from parameter-space correlators, 2021. https://arxiv.org/abs/2106.00694

    Pith/arXiv arXiv 2021

  46. [46]

    Halverson, A

    J. Halverson, A. Maiti and K. Stoner,Neural networks and quantum field theory, 2021. https://doi.org/10.1088/2632-2153/abeca3, https://arxiv.org/abs/2008.08601

    Pith/arXiv arXiv doi:10.1088/2632-2153/abeca3 2021

  47. [47]

    Halverson,Building quantum field theories out of neurons,

    J. Halverson,Building quantum field theories out of neurons,

  48. [48]

    https://arxiv.org/abs/2112.04527

    Pith/arXiv arXiv

  49. [49]

    Jacot, F

    A. Jacot, F. Gabriel and C. Hongler,Neural tangent kernel: Convergence and generalization in neural networks, 2020. https://arxiv.org/abs/1806.07572. Pre-Strings Lectures on Artificial Intelligence 45

    Pith/arXiv arXiv 2020

  50. [50]

    J. Lee, L. Xiao, S. S. Schoenholz, Y. Bahri, R. Novak, J. Sohl-Dickstein et al.,Wide neural networks of any depth evolve as linear models under gradient descent, 2019. https://doi.org/10.1088/1742-5468/abc62b, https://arxiv.org/abs/1902.06720

    Pith/arXiv arXiv doi:10.1088/1742-5468/abc62b 2019

  51. [51]

    Pehlevan and B

    C. Pehlevan and B. Bordelon,Lecture notes on infinite-width limits of neural networks, August, 2024. https://mlschool.princeton.edu/events/2023/pehlevan

    2024

  52. [52]

    Yang and E

    G. Yang and E. J. Hu,Feature learning in infinite-width neural networks, 2022. https://arxiv.org/abs/2011.14522

    Pith/arXiv arXiv 2022

  53. [53]

    Bordelon and C

    B. Bordelon and C. Pehlevan,Self-consistent dynamical field theory of kernel evolution in wide neural networks, 2022. https://arxiv.org/abs/2205.09653

    Pith/arXiv arXiv 2022

  54. [54]

    D. A. Roberts, S. Yaida and B. Hanin,The principles of deep learning theory, 2021. https://doi.org/10.1017/9781009023405, https://arxiv.org/abs/2106.10165

    Pith/arXiv arXiv doi:10.1017/9781009023405 2021

  55. [55]

    Yaida,Meta-principled family of hyperparameter scaling strategies, 2022

    S. Yaida,Meta-principled family of hyperparameter scaling strategies, 2022. https://arxiv.org/abs/2210.04909

    Pith/arXiv arXiv 2022

  56. [56]

    S. S. Schoenholz, J. Gilmer, S. Ganguli and J. Sohl-Dickstein, Deep information propagation, 2017. https://arxiv.org/abs/1611.01232

    Pith/arXiv arXiv 2017

  57. [57]

    Frank, J

    S. Frank, J. Halverson, A. Maiti and F. Ruehle,Fermions and supersymmetry in neural network field theories, 2025. https://arxiv.org/abs/2511.16741

    arXiv 2025

  58. [58]

    Osterwalder and R

    K. Osterwalder and R. Schrader,Axioms for Euclidean Green’s Functions,Commun. Math. Phys.31(1973) 83

    1973

  59. [59]

    Simmons-Duffin,TASI lectures on the conformal bootstrap,

    D. Simmons-Duffin,TASI lectures on the conformal bootstrap,

  60. [60]

    https://arxiv.org/abs/1602.07982

    Pith/arXiv arXiv

  61. [61]

    Ferko and J

    C. Ferko and J. Halverson,Quantum mechanics and neural networks, 2025. https://arxiv.org/abs/2504.05462

    Pith/arXiv arXiv 2025

  62. [62]

    Halverson, J

    J. Halverson, J. Naskar and J. Tian,Conformal fields from neural networks, 2025. https://doi.org/10.1007/JHEP10(2025)039, https://arxiv.org/abs/2409.12222

    work page doi:10.1007/jhep10(2025)039 2025

  63. [63]

    Robinson,Virasoro symmetry in neural network field theories, 2026

    B. Robinson,Virasoro symmetry in neural network field theories, 2026. https://arxiv.org/abs/2512.24420

    arXiv 2026

  64. [64]

    Capuozzo, B

    P. Capuozzo, B. Robinson and B. Suzzoni,Conformal defects in neural network field theories, 2026. https://doi.org/10.1007/JHEP05(2026)124, https://arxiv.org/abs/2512.07946

    Pith/arXiv arXiv doi:10.1007/jhep05(2026)124 2026

  65. [65]

    Ferko, S

    C. Ferko, S. Frank, J. Halverson and V. Jejjala,Anomalies in neural network field theory, 2026. https://arxiv.org/abs/2605.12488

    Pith/arXiv arXiv 2026

  66. [66]

    David, A

    F. David, A. Kupiainen, R. Rhodes and V. Vargas,Liouville quantum gravity on the riemann sphere, 2015. https://arxiv.org/abs/1410.7318

    Pith/arXiv arXiv 2015

  67. [67]

    Probabilistic paths to Quantum Field Theory

    Simons Collaboration on Probabilistic Paths to Quantum Field Theory, “Probabilistic paths to Quantum Field Theory.”https://probabilistic-qft.org/, 2026

    2026

  68. [68]

    Jiang, B

    M. Jiang, B. P. Czajka, E. B. James, A. J. Hardaway, C. D. Achammer, Q. Su et al.,Neural network approach to dirac quantum field theory,Phys. Rev. A112(2025) 062223

    2025

  69. [69]

    J. N. Howard, M. S. Klinger, A. Maiti and A. G. Stapleton, Bayesian rg flow in neural network field theories, 2025. https://doi.org/10.21468/SciPostPhysCore.8.1.027, https://arxiv.org/abs/2405.17538

    Pith/arXiv arXiv doi:10.21468/scipostphyscore.8.1.027 2025

  70. [70]

    D. S. Ageev and Y. A. Ageeva,Neural network quantum field theory from transformer architectures, 2026. https://arxiv.org/abs/2602.10209. Pre-Strings Lectures on Artificial Intelligence 46

    arXiv 2026

  71. [71]

    Zhang,Optimal architecture and fundamental bounds in neural network field theory, 2026

    Z. Zhang,Optimal architecture and fundamental bounds in neural network field theory, 2026. https://arxiv.org/abs/2604.27050

    Pith/arXiv arXiv 2026

  72. [72]

    Ferko, J

    C. Ferko, J. Halverson and A. Mutchler,Universality of neural network field theory, 2026. https://arxiv.org/abs/2601.14453

    arXiv 2026

  73. [73]

    Frank and J

    S. Frank and J. Halverson,String theory from infinite width neural networks, 2026. https://arxiv.org/abs/2601.06249

    arXiv 2026

  74. [74]

    Ferko, J

    C. Ferko, J. Halverson, V. Jejjala and B. Robinson, Topological effects in neural network field theory, 2026. https://arxiv.org/abs/2604.02313

    Pith/arXiv arXiv 2026

  75. [75]

    Dorn and H

    H. Dorn and H. J. Otto,Two and three-point functions in liouville theory, 1994. https://doi.org/10.1016/0550-3213(94)00352-1, https://arxiv.org/abs/hep-th/9403141

    Pith/arXiv arXiv doi:10.1016/0550-3213(94)00352-1 1994

  76. [76]

    A. B. Zamolodchikov and A. B. Zamolodchikov,Structure constants and conformal bootstrap in liouville field theory,

  77. [77]

    https://doi.org/10.1016/0550-3213(96)00351-3, https://arxiv.org/abs/hep-th/9506136

    Pith/arXiv arXiv doi:10.1016/0550-3213(96)00351-3

  78. [78]

    Chatterjee and E

    S. Chatterjee and E. Witten,Liouville theory: An introduction to rigorous approaches, 2024. https://arxiv.org/abs/2404.02001

    Pith/arXiv arXiv 2024

  79. [79]

    Song and S

    Y. Song and S. Ermon,Generative modeling by estimating gradients of the data distribution, 2020. https://arxiv.org/abs/1907.05600

    Pith/arXiv arXiv 2020

  80. [80]

    Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon and B. Poole,Score-based generative modeling through stochastic differential equations, 2021. https://arxiv.org/abs/2011.13456

    Pith/arXiv arXiv 2021

Showing first 80 references.