Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

Inductive bias, not physics: what neural renormalization-group and machine-learned holographic maps can and cannot identify

Yan Ru Pei

Published Nov 7, 2026• 0 citations
Publisher

Abstract

Neural renormalization-group (RG) methods and machine-learned holographic maps train a network on lattice samples or on a lattice action and read physics off the result: coarse couplings, critical exponents, or an emergent bulk geometry. We ask which of these outputs the training objective and the readout themselves determine. For a bijection trained by a divergence to a factorized target with a free coarse coupling, every admissible coupling attains zero loss through a smooth optimal-transport map that respects every symmetry acting identically on field and latents, so neither more training nor an equivariant architecture selects the coupling. Restricting the class can: in the Gaussian chain a zero-loss linear map that is local below the coarse correlation length has the decimation coupling, the threshold is sharp, and on the square lattice no finite-range linear map reaches zero loss. A Gaussian kernel on a network graph yields a distance bounded above and below by the graph distance whatever the data, and exact wavelet latents of a power-law Gaussian give an isotropic hyperbolic geometry except along chains through the wavelet's jumps, which a free-fermion computation ties to the anisotropy reported for the exact holographic mapping. We then re-examine three published pipelines. A hyperbolicity test of neural-RG holography is passed by kernels built without boundary data. In RGFlow at the reported sizes, general linear maps reach zero loss at a continuum of couplings, whereas in a small, unconverged φ⁴ computation RGFlow's own RealNVP class selects a coupling through its Jacobian structure, close to the decimation value in the Gaussian sector. In inverse RG of the three-dimensional spin glass, with a ten-parameter stand-in upscaler on small lattices, the one clean upscaler that closes under blocking reproduced the forward Monte Carlo RG fixed point of its training pair, non-closing ones drifted, and the lower training loss went to a non-closing upscaler in both training pairs. The sharp results are Gaussian. We close with diagnostics, each with a recipe.

DOI
10.5281/zenodo.24042561

Related papers

Scaling and Renormalization in Neural Networks: An Effective Renormalization-Group Description of Grokking

Min Chen, Junqi Wang, Zeguan Wu +2 more

Grokking is a delayed transition from memorization to generalization in deep neural networks (NNs). We examine whether scaling and renormalization group (RG) theories can describe this dynamics, establishing a quantitative physical framework for neural generalization transitions. For modular addition on a two-layer NN, an ensemble of trained models defines an order parameter (the squared weight norm) and fluctuation proxies for the heat capacity and the susceptibility, with training epoch and weight decay serving as thermal and field control coordinates. To bridge neural architectures to real-space RG, we construct an exact coupling matrix via a second-order expansion of the loss with respect to per-neuron norms and use its Fiedler vector to define an effective geometry. This spectral ordering yields a monotonic spatial decay of coupling strength that is absent under random permutations. We then execute real-space RG by block-spin aggregating adjacent neurons and parameterizing the coupling matrix into off-diagonal band strengths. Linearizing this local RG map yields relevant scaling exponents and five derived critical exponents, with direct empirical estimates matching RG predictions within 1-10%. Overall, our framework provides a consistent effective description of grokking, suggesting the scaling and renormalization group formalism as a tool to investigate NN dynamics.

cs.LGCross-listedarXiv:2610.36113•Oct 28, 2026

Neural Network Renormalization Group

Shuo-Hui Li, Lei Wang

We present a variational renormalization group (RG) approach based on a reversible generative model with hierarchical architecture. The model performs hierarchical change-of-variables transformations from the physical space to a latent space with reduced mutual information. Conversely, the neural network directly maps independent Gaussian noises to physical configurations following the inverse RG flow. The model has an exact and tractable likelihood, which allows unbiased training and direct access to the renormalized energy function of the latent variables. To train the model, we employ probability density distillation for the bare energy function of the physical problem, in which the training loss provides a variational upper bound of the physical free energy. We demonstrate practical usage of the approach by identifying mutually independent collective variables of the Ising model and performing accelerated hybrid Monte Carlo sampling in the latent space. Lastly, we comment on the connection of the present approach to the wavelet formulation of RG and the modern pursuit of information preserving RG.

arXiv:1802.02840•Physical Review Letters•Dec 26, 2018•160 citations

Holography as deep learning

Wen-Cong Gan, Fu-Wen Shu

Quantum many-body problem with exponentially large degrees of freedom can be reduced to a tractable computational form by neural network method [G. Carleo and M. Troyer, Science 355 (2017) 602, arXiv:1606.02318.] The power of deep neural network (DNN) based on deep learning is clarified by mapping it to renormalization group (RG), which may shed lights on holographic principle by identifying a sequence of RG transformations to the AdS geometry. In this paper, we show that any network which reflects RG process has intrinsic hyperbolic geometry, and discuss the structure of entanglement encoded in the graph of DNN. We find the entanglement structure of DNN is of Ryu–Takayanagi form. Based on these facts, we argue that the emergence of holographic gravitational theory is related to deep learning process of the quantum-field theory.

arXiv:1705.05750•International Journal of Modern Physics D•Oct 1, 2017•57 citations

Application of deep neural networks for computing the renormalization group flow of the two-dimensional ϕ 4 field theory

Yueqi Zhao, M. M. Fogler, Yi‐Zhuang You

Abstract We introduce RGFlow, a deep neural network–based real-space renormalization group (RG) framework tailored for continuum scalar field theories. Leveraging generative capabilities of flow-based neural networks, RGFlow autonomously learns real-space RG transformations from data without prior knowledge of the underlying model. In contrast to conventional approaches, RGFlow is bijective (information-preserving) and is optimized based on the principle of minimal mutual information. We demonstrate the method on two examples. The first one is a one-dimensional Gaussian model, where RGFlow is shown to learn the classical decimation rule. The second is the two-dimensional φ 4 theory, where the network successfully identifies a Wilson–Fisher-like critical point and provides an estimate of the correlation-length critical exponent.

arXiv:2510.06508•Machine Learning Science and Technology•Feb 23, 2026•1 citation

Renormalization Group connected to Neural Networks

J.J. ter Hoeve

Even though deep learning has proved to be very powerful as the core method of machine learning, theoretical understanding behind its success is still unclear. It is been pointed out in recent years that the behaviour of deep neural networks is reminiscent of a fundamental framework in statistical physics: the renormalization group (RG). Motivated by this analogy, we develop an analytical method of directly obtaining the trained weights W resulting from a training set of 1D-Ising samples with coupling J and temperature T. The found relation W(J) drives a flow that takes 1D-Ising configurations at non-zero temperature to the critical point at T = 0. This behaviour is opposite to what a typical RG-flow dictates.

•Jan 1, 2018•0 citations

Machine Learning the Strong Disorder Renormalization Group Method for Disordered Quantum Spin Chains

A. Ustyuzhanin, J. Vahedi, S. Kettemann

We train machine learning algorithms to infer the entanglement structure of disordered long-range interacting quantum spin chains by learning from the strong disorder renormalisation group (SDRG) method. The system consists of $S=1/2$-quantum spins coupled by antiferromagnetic power-law interactions with decay exponent $α$ at random positions on a one-dimensional chain. Using SDRG as a physics-informed teacher, we compare a Random Forest classifier as a classical baseline with a graph neural network (GNN) that operates directly on the interaction graph and learns a bond-ranking rule mirroring the SDRG decimation policy. The GNN achieves a disorder-averaged pairing accuracy close to one and reproduces the entanglement entropy $S(\ell)$ in excellent quantitative agreement with SDRG across all subsystem sizes and interaction exponents. RG flow heat maps confirm that the GNN learns the sequential decimation hierarchy rather than merely fitting final-state observables. Finite-temperature entanglement properties are incorporated via the SDRGX framework through a two-stage strategy, using the zero-temperature GNN to generate the RG flow and sampling thermal occupations from the canonical ensemble, yielding results in agreement with both numerical SDRGX and analytical predictions without retraining.

cond-mat.dis-nnCross-listedarXiv:2603.05164•Machine Learning Science and Technology•Mar 6, 2026•1 citation
Neural renormalization-group (RG) methods and machine-learned holographic maps train a network on lattice samples or on a lattice action and read physics off the result: coarse couplings, critical exponents, or an emergent bulk geometry. We ask which of these outputs the training objective and the readout themselves determine. For a bijection trained by a divergence to a factorized target with a free coarse coupling, every admissible coupling attains zero loss through a smooth optimal-transport map that respects every symmetry acting identically on field and latents, so neither more training nor an equivariant architecture selects the coupling. Restricting the class can: in the Gaussian chain a zero-loss linear map that is local below the coarse correlation length has the decimation coupling, the threshold is sharp, and on the square lattice no finite-range linear map reaches zero loss. A Gaussian kernel on a network graph yields a distance bounded above and below by the graph distance whatever the data, and exact wavelet latents of a power-law Gaussian give an isotropic hyperbolic geometry except along chains through the wavelet's jumps, which a free-fermion computation ties to the anisotropy reported for the exact holographic mapping. We then re-examine three published pipelines. A hyperbolicity test of neural-RG holography is passed by kernels built without boundary data. In RGFlow at the reported sizes, general linear maps reach zero loss at a continuum of couplings, whereas in a small, unconverged φ⁴ computation RGFlow's own RealNVP class selects a coupling through its Jacobian structure, close to the decimation value in the Gaussian sector. In inverse RG of the three-dimensional spin glass, with a ten-parameter stand-in upscaler on small lattices, the one clean upscaler that closes under blocking reproduced the forward Monte Carlo RG fixed point of its training pair, non-closing ones drifted, and the lower training loss went to a non-closing upscaler in both training pairs. The sharp results are Gaussian. We close with diagnostics, each with a recipe.
Grokking is a delayed transition from memorization to generalization in deep neural networks (NNs). We examine whether scaling and renormalization group (RG) theories can describe this dynamics, establishing a quantitative physical framework for neural generalization transitions. For modular addition on a two-layer NN, an ensemble of trained models defines an order parameter (the squared weight norm) and fluctuation proxies for the heat capacity and the susceptibility, with training epoch and weight decay serving as thermal and field control coordinates. To bridge neural architectures to real-space RG, we construct an exact coupling matrix via a second-order expansion of the loss with respect to per-neuron norms and use its Fiedler vector to define an effective geometry. This spectral ordering yields a monotonic spatial decay of coupling strength that is absent under random permutations. We then execute real-space RG by block-spin aggregating adjacent neurons and parameterizing the coupling matrix into off-diagonal band strengths. Linearizing this local RG map yields relevant scaling exponents and five derived critical exponents, with direct empirical estimates matching RG predictions within 1-10%. Overall, our framework provides a consistent effective description of grokking, suggesting the scaling and renormalization group formalism as a tool to investigate NN dynamics.
We present a variational renormalization group (RG) approach based on a reversible generative model with hierarchical architecture. The model performs hierarchical change-of-variables transformations from the physical space to a latent space with reduced mutual information. Conversely, the neural network directly maps independent Gaussian noises to physical configurations following the inverse RG flow. The model has an exact and tractable likelihood, which allows unbiased training and direct access to the renormalized energy function of the latent variables. To train the model, we employ probability density distillation for the bare energy function of the physical problem, in which the training loss provides a variational upper bound of the physical free energy. We demonstrate practical usage of the approach by identifying mutually independent collective variables of the Ising model and performing accelerated hybrid Monte Carlo sampling in the latent space. Lastly, we comment on the connection of the present approach to the wavelet formulation of RG and the modern pursuit of information preserving RG.
Quantum many-body problem with exponentially large degrees of freedom can be reduced to a tractable computational form by neural network method [G. Carleo and M. Troyer, Science 355 (2017) 602, arXiv:1606.02318.] The power of deep neural network (DNN) based on deep learning is clarified by mapping it to renormalization group (RG), which may shed lights on holographic principle by identifying a sequence of RG transformations to the AdS geometry. In this paper, we show that any network which reflects RG process has intrinsic hyperbolic geometry, and discuss the structure of entanglement encoded in the graph of DNN. We find the entanglement structure of DNN is of Ryu–Takayanagi form. Based on these facts, we argue that the emergence of holographic gravitational theory is related to deep learning process of the quantum-field theory.
Abstract We introduce RGFlow, a deep neural network–based real-space renormalization group (RG) framework tailored for continuum scalar field theories. Leveraging generative capabilities of flow-based neural networks, RGFlow autonomously learns real-space RG transformations from data without prior knowledge of the underlying model. In contrast to conventional approaches, RGFlow is bijective (information-preserving) and is optimized based on the principle of minimal mutual information. We demonstrate the method on two examples. The first one is a one-dimensional Gaussian model, where RGFlow is shown to learn the classical decimation rule. The second is the two-dimensional φ 4 theory, where the network successfully identifies a Wilson–Fisher-like critical point and provides an estimate of the correlation-length critical exponent.
Even though deep learning has proved to be very powerful as the core method of machine learning, theoretical understanding behind its success is still unclear. It is been pointed out in recent years that the behaviour of deep neural networks is reminiscent of a fundamental framework in statistical physics: the renormalization group (RG). Motivated by this analogy, we develop an analytical method of directly obtaining the trained weights W resulting from a training set of 1D-Ising samples with coupling J and temperature T. The found relation W(J) drives a flow that takes 1D-Ising configurations at non-zero temperature to the critical point at T = 0. This behaviour is opposite to what a typical RG-flow dictates.
We train machine learning algorithms to infer the entanglement structure of disordered long-range interacting quantum spin chains by learning from the strong disorder renormalisation group (SDRG) method. The system consists of S=1/2S=1/2S=1/2-quantum spins coupled by antiferromagnetic power-law interactions with decay exponent ααα at random positions on a one-dimensional chain. Using SDRG as a physics-informed teacher, we compare a Random Forest classifier as a classical baseline with a graph neural network (GNN) that operates directly on the interaction graph and learns a bond-ranking rule mirroring the SDRG decimation policy. The GNN achieves a disorder-averaged pairing accuracy close to one and reproduces the entanglement entropy S(ℓ)S(\ell)S(ℓ) in excellent quantitative agreement with SDRG across all subsystem sizes and interaction exponents. RG flow heat maps confirm that the GNN learns the sequential decimation hierarchy rather than merely fitting final-state observables. Finite-temperature entanglement properties are incorporated via the SDRGX framework through a two-stage strategy, using the zero-temperature GNN to generate the RG flow and sampling thermal occupations from the canonical ensemble, yielding results in agreement with both numerical SDRGX and analytical predictions without retraining.