Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

Adaptive kernel predictors from feature-learning infinite limits of neural networks

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read Infinitely wide feature-learning neural networks are kernel machines with data-dependent kernels.

desk verdict A serious theory paper with a genuinely new aNBK construction and a real gap in the aNTK claim: Appendix F.2 concedes the fixed-point equations don't determine the kernel. read the letter →

arxiv 2502.07998 v2 pith:JCMSLQ5G submitted 2025-02-11 cs.LG cond-mat.dis-nnstat.ML

classification cs.LGcond-mat.dis-nnstat.ML MSC 68T0782C32
keywords kernelmachinesfeaturelearninginfinite-widthlimitsneuraltangentBayesiannetworksdynamicalmeanfieldtheorymaximalupdateparameterizationadaptivekernels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that the rich, feature-learning infinite-width limit of neural networks—not just the lazy limit—is described by kernel machines, only now with a kernel that adapts to the training data. It derives two explicit predictors: the adaptive Neural Bayesian Kernel (aNBK), from the infinite-width posterior of noisy gradient flow, and the adaptive Neural Tangent Kernel (aNTK), from the fixed points of gradient flow with weight decay. For squared loss and a linear readout each predictor is ridge regression with a data-dependent kernel, $\Phi_L$ or $K_{\mathrm{aNTK}}$, whose entries are fixed by saddle-point or dynamical mean-field theory (DMFT) fixed-point equations rather than by the initial weights. If correct, infinitely wide feature-learning MLPs and CNNs can be evaluated and understood as kernel regressors, and the gap between rich and lazy training is captured by kernel adaptation.

What carries the argument

The carrying objects are two data-dependent kernel matrices and the equations that determine them. For aNBK, the layer feature kernels $\Phi^\ell_{\mu\nu}=N^{-1}\phi(h^\ell_\mu)\cdot\phi(h^\ell_\nu)$ and their conjugate dual kernels $\widehat{\Phi}^\ell$ are fixed by the saddle point of the Bayesian action $S=-\frac12\sum_{\ell=1}^L\mathrm{Tr}(\Phi^\ell\widehat{\Phi}^\ell)+\frac{\gamma_0^2}{2}y^\top(\beta^{-1}I+\Phi^L/\lambda_L)^{-1}y-\sum_{\ell=1}^{L-1}\ln Z^\ell[\Phi^{\ell-1},\widehat{\Phi}^\ell]$, a min-max problem whose solution is the kernel in Eq. (5). For aNTK, the machinery is the DMFT field dynamics for pre-activations $h^\ell_\mu(t)$ and pre-gradient signals $z^\ell_\mu(t)$; with weight decay the initial-condition noise is exponentially suppressed, and the infinite-time fixed point of these fields defines $K_{\mathrm{aNTK}}=\lim_{t\to\infty}\sum_{\ell}G^{\ell+1}_{\mu\nu}(t,t)\Phi^\ell_{\mu\nu}(t,t)$ in Eq. (11). These objects are what make the infinite-width limit a kernel machine: they convert the training data and task labels into a kernel without perturbing $\gamma_0$.

What would settle it

Train two very wide two-layer ReLU networks with the same data, loss, and hyperparameters but different initial weight distributions (Gaussian versus Laplace) in the rich regime with weight decay; if the converged predictors differ at large width, then $K_{\mathrm{aNTK}}$ depends on training history rather than being a deterministic data-dependent kernel.

Watch

Extended reading notes

Core claim

The central discovery is that infinitely wide MLPs and CNNs trained in the rich, feature-learning regime (maximal-update parameterization, $\gamma_0=\Theta(1)$) converge to kernel regressors with data-dependent kernels. In the Bayesian setting, taking $t\to\infty$ at fixed temperature and then $N\to\infty$ gives $f_{\mathrm{aNBK}}(x)=\Phi_L(x)^\top(\Phi_L+\lambda_L I/\beta)^{-1}y$, where $\Phi_L$ is the converged last-layer feature kernel obtained from the min-max saddle point of the action $S(\{\Phi^\ell,\widehat{\Phi}^\ell\})$ in Eq. (6). In the gradient-flow setting, taking $N\to\infty$ and then $\beta\to\infty$ with weight decay $\lambda>0$ gives $f_{\mathrm{aNTK}}(x)=k_{\mathrm{aNTK}}(x)^\top(K_{\mathrm{aNTK}}+\lambda L\kappa I)^{-1}y$, where $K_{\mathrm{aNTK}}$ is the infinite-time limit of the DMFT kernel dynamics. These kernels are deterministic functions of the data and hyperparameters, recoverable numerically without expanding in the richness parameter $\gamma_0$; when $\gamma_0\to0$ they reduce to the NNGP and NTK predictors. The paper also solves the deep linear case exactly, obtaining layerwise kernel-label overlap recursions and scaling laws, and shows numerically that the adaptive kernels match finite-width trained networks and achieve lower test loss than lazy kernels on CIFAR10 and MNIST.

Load-bearing premise

The argument assumes that infinitely wide training settles into a unique final state whose kernel is purely a function of the data and training settings; the appendix notes the equations fix only the first two moments of the neuron's pre-activation distribution, so the full distribution—and hence the kernel—may still depend on how training started.

Editorial extensions

If this is right

  • At $\gamma_0\to0$, the aNBK and aNTK predictors reduce to the NNGP and NTK kernels, so lazy kernel theory is the weak-feature-learning edge of a single adaptive-kernel family.
  • Feature learning at infinite width produces non-Gaussian pre-activation densities and label-aligned layer kernels; these are directly computable from the saddle-point or DMFT solutions, so kernel regression with adaptive kernels inherits them.
  • The aNBK and aNTK predictors can be evaluated numerically (Algorithms 1 and 2) and match finite-width trained networks at widths such as $N=5000$ on CIFAR10, making them candidate stand-ins for actual rich training.
  • For deep linear networks with whitened data, the kernel-label overlaps $c_\ell$ grow exponentially in layer depth, so alignment with labels propagates from the last layer backward; the scaling laws $c_L\sim 1+L\gamma_0^2$, $c_L\sim\gamma_0^{2L/(L+1)}$, and $c_L\sim L\gamma_0^2/\ln(L\gamma_0^2)$ characterize the lazy, rich-shallow, and deep limits.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper does not say this, but if the aNTK claim holds, weight decay is not just regularization: it is the mechanism that erases initialization memory and makes the infinite-width limit a history-independent kernel machine, which suggests that tuning $\lambda$ controls how kernel-like a trained network is and that at $\lambda=0$ no kernel representation of the converged predictor should be expec
  • Because Appendix F.2 reports that the DMFT fixed-point equations fix only the first two moments of the single-site density and the full marginal remains history dependent, the aNTK kernel defined in Eq. (11) may not be unique unless additional conditions select a density; an extension would be to identify those extra fixed-point constraints or characterize the basin of densities.
  • One testable extension is to use the adaptive kernels as cheap surrogates for feature-learning networks in scaling-law studies: if the kernel predictor matches finite-width training at moderate $N$, it could replace expensive runs when predicting scaling exponents.
  • The min-max solver's convergence difficulty for $L>1$ (helped by warm starts in Fig. 4) suggests that proving convergence or finding a convex reformulation of Eq. (8) would be needed before aNBK can serve as a drop-in training surrogate for deep architectures.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper studies infinite-width limits of neural networks in the feature-learning (maximal-update) regime and claims that, in two settings, the learned predictor is a kernel machine with a data-dependent kernel. The first setting is infinite-time noisy gradient flow at fixed temperature: the authors derive a statistical-mechanics action whose saddle point leads to a min-max problem for layer feature kernels and dual kernels, yielding the aNBK predictor (Eq. 5). The second setting is infinite-time gradient flow with weight decay: using DMFT equations imported from Bordelon and Pehlevan (2022), the authors define the aNTK predictor (Eq. 11) in terms of fixed-point feature and gradient kernels. The paper solves the deep linear case exactly, provides numerical algorithms for the nonlinear case, and compares the resulting predictors against finite-width MLP and CNN simulations, reporting that the adaptive kernels outperform NNGP and NTK predictors.

Significance. If the central claims hold, this is a significant conceptual advance: it extends the kernel-machine description of infinite-width neural networks beyond the lazy regime, giving explicit data-dependent kernels that can in principle be computed without simulating finite-width networks. The paper's strengths are its explicit predictor formulas, the exact solution and scaling laws for deep linear networks, the concrete numerical algorithms, and the careful finite-width comparisons. The main open gap, namely whether the aNTK fixed-point equations uniquely determine the kernel, does not diminish the interest of the results, but it must be resolved or clearly scoped before the abstract's claim is justified.

major comments (3)
  1. [Appendix F.2 / Eq. (11)] The aNTK representer theorem is not yet established because the fixed-point equations of the DMFT do not, in general, determine the kernel. Appendix F.2 states that the constraints (81) fix only the first two moments of the single-site pre-activation density and that the full marginal 'remains history dependent'; one 'has to track the entire update dynamics in order to get the full description.' For nonlinear activations, K_aNTK in Eq. (12) averages nonlinear functionals of this density, such as <phi(h)phi(h)> and <phi'(h)z phi'(h)z>, so knowledge of the first two moments is insufficient to determine K_aNTK. The F.2 linear example does not resolve this, since for linear activations the kernel depends only on the second moment. The authors should either prove that the invariant measure, or at least the kernel averages, is unique for the activations and architectures considered, or demonstrate numerically that K_aNTK is independent of the initialization and training trajectory.
  2. [§3.1, Eq. (8); Algorithm 1] The aNBK derivation replaces the N-to-infinity partition function by the saddle point of the action S in Eq. (8) and identifies this with a global min-max problem, but no argument is given for uniqueness, convexity-concavity, or the dominance of a global saddle point. This is not a purely formal point: the numerical solver is reported in Section 5 and Appendix G.2 to converge with difficulty for L>1, and Figure 7 was produced by initializing the solver with the empirical kernels obtained from Langevin simulations. Such a warm start can select a nearby local fixed point, so the reported agreement between theory and experiment does not yet certify Eq. (8) as a data-only predictor. Please report convergence from lazy or random initializations for the comparisons, or explicitly restrict the claims to locally stable saddle points.
  3. [§3.2, Eq. (13) and Eq. (14)] The derivation of Eq. (11) treats the final K_aNTK as a fixed matrix at the DMFT fixed point, but Eq. (13) shows that the feature and gradient kernels evolve along a non-Markovian trajectory whose integral terms depend on the full history of the error signal. Exponential suppression of the initial-condition terms addresses only one source of history dependence; Appendix F.2 shows that another source persists even at the fixed point. A closure argument is needed to show that the fixed-point conditions (81), together with the predictor equation, select a unique K_aNTK for each dataset and hyperparameter setting. Without such an argument, Eq. (11) is a trajectory-dependent description rather than a representer theorem with a data-dependent kernel fixed by the DMFT fixed point.
minor comments (5)
  1. [Eq. (13)] The notation ẍφ(h^ℓ_ν(t′)) is used without defining whether the derivative is with respect to the argument or with respect to time; please clarify.
  2. [Fig. 3] The caption lists the green curve as NNGPK and the orange and red curves as aNTK and aNBK, but the legend colors in the figure are not self-evident from the text; please harmonize the caption with the actual color coding.
  3. [Table 1] The table of limiting orders is difficult to read because the column entries are compressed; please reformat so that the order of limits for NNGPK, aNBK, NTK, and aNTK is unambiguous.
  4. [Appendix G.6] The complexity comparison between the theory solver and finite-width training is asymptotic only; for the reported values N=1024, P up to 1000, and T=20000, wall-clock times would be more informative than the statement that the theory solver is 'typically much lower' in cost.
  5. [General] The paper would benefit from a code/data availability statement; the hyperparameter details in Appendix G are helpful but not sufficient for exact reproduction of the kernel alignment numbers.

Circularity Check

1 steps flagged · score 3.0 of 10

Localized circularity in Fig. 7 warm start; central kernel-predictor derivations are self-contained, with the aNTK fixed-point gap flagged as a non-circular correctness risk.

  1. fitted input called prediction [Appendix A.1 (warm-start paragraph for Fig. 7)]
    "However, an easier strategy for convergence is to initialize the solver Algorithm 1 with the empirical NN kernels{Φℓ}Lℓ=1 at convergence obtained from Langevin simulations and perturbed with a multiplicative Gaussian noise. In this way, one only needs to solve for the dual variables. This warm start allows faster convergence. This is what we did to produce Fig. 7."

    The theoretical curves labeled 'theory' in Fig. 7 are generated by seeding Algorithm 1 with the empirical per-layer kernels Φℓ measured from the very Langevin simulations they are meant to predict, perturbed only by multiplicative Gaussian noise, and then solving only for the dual variables Φ̂ℓ. Since the reported quantities—the kernel-label alignments A(Φℓ, yy^T) and the layer kernels themselves—are dominated by the initialized Φℓ, the agreement between theory and experiment in Fig. 7 is partly true by construction. The solver is not independently predicting the learned kernels; it is returning a nearby fixed point of the action reached from the empirical kernels, so this specific 'prediction' reduces to the empirical input plus a local relaxation.

full rationale

The paper's central claims—the aNBK predictor of Eq. (5) from the saddle point of the Bayesian posterior and the aNTK predictor of Eq. (11) from DMFT fixed points—are derivations rather than definitions: the hyperparameters {γ0, β, λ} are inputs, and the kernels are solved from self-consistent equations, so the main derivation chain is not circular. The one concrete circular step is in Appendix A.1: to produce Fig. 7, the aNBK solver is initialized with the empirical layer kernels from the Langevin simulations that the theory is supposed to reproduce. With Φℓ fixed (up to multiplicative noise) at the empirical values, the solver only adjusts the dual variables, so the reported agreement of kernel alignments and layer kernels is substantially inherited from the initialization rather than independently predicted. This is a fitted-input-called-prediction pattern, though it is confined to one numerical comparison and does not invalidate the analytic derivation by itself. Separately, the aNTK fixed-point underdetermination admitted in Appendix F.2 is a real load-bearing gap: the constraints fix only the first two moments, while the kernel averages in Eq. (12) require the full joint density, so K_aNTK may be history-dependent and not a data-only kernel. I do not count this as circularity because it is a missing uniqueness or closure argument, not an equivalence of the predictor to its inputs by construction. The DMFT citation to Bordelon and Pehlevan (2022) is same-group but is prior derivational work, not a self-citation used to forbid alternatives, so it does not raise the circularity score.

Assumptions & free parameters 3 free parameters · 7 assumptions · 3 invented entities

The central claims rest on a small set of hand-set hyperparameters (gamma0, lambda, beta) that are inputs of the theory, not fitted to the predicted outputs; the deep-linear recursion is parameter-free. The main assumptions are standard statistical-mechanics tools (saddle point/Laplace evaluation, DMFT self-consistency, Langevin posterior sampling, CLT for preactivations at P = Theta(1)) plus modelling restrictions (mu-P scaling, homogeneous activations for aNTK, non-saturated readout). The paper introduces the adaptive kernels PhiL and K_aNTK and the dual order parameters hatPhi_l; the two kernels carry falsifiable content and are tested against finite-width simulations, while the dual kernels are internal bookkeeping. Appendix F.2 partially undermines the abstraction of the aNTK fixed point as history independent.

free parameters (3)
  • gamma0 (feature-learning strength) = 0.3 for Fig. 3; 0.2-4.0 for Figs. 10-11; 0.5-10 for Fig. 2
    Order-one parameter of the mu-P scaling controlling the rich/lazy interpolation; hand-set per experiment, not fitted to the predicted test losses or densities.
  • lambda (weight decay / ridge) = 1.0 for MLPs; 0.01 for CNN comparisons
    Regularization hyperparameter entering the kernels as lambdaL (Eq. 5) and lambdaL kappa (Eq. 11); hand-set.
  • beta (inverse temperature) = infinity for aNTK; 50 for aNBK Langevin experiments
    Temperature of the noisy dynamics and of the posterior (Eq. 9); hand-set; the beta->infinity limit is used for aNTK.
assumptions (7)
  • standard math Saddle point / Laplace evaluation of the Bayesian action S at N->infinity
    Eq. (24): Z = integral dPhi dhatPhi exp(-N S) is dominated by stationary points; used to justify Eq. (8). No proof of saddle uniqueness or of the min-max signature.
  • domain assumption DMFT field equations (13) exactly describe infinite-width gradient flow dynamics
    Imported from Bordelon and Pehlevan (2022); requires neuron decoupling and kernel concentration as N->infinity; underlies the aNTK derivation (Section 3.2, App. E).
  • domain assumption Langevin dynamics (3) samples the Gibbs posterior (9) at stationarity
    Needed for the aNBK interpretation of the t->infinity limit at fixed beta; cites Welling and Teh (2011) and Mingard et al. (2020); assumes ergodicity and correct isotropic noise.
  • domain assumption Activation homogeneity (degree kappa) for the aNTK weight-decay fixed point
    The ridge lambdaL kappa in Eqs. (10)-(11) requires f(a theta) = a^kappa f(theta); holds for ReLU, excludes tanh in the aNTK setting (Section 3.2, App. E).
  • domain assumption sigma'(s_mu) != 0 for all patterns in the aNTK derivation
    Footnote 3: saturated readouts block gradient propagation and prevent convergence to a kernel predictor.
  • domain assumption P = Theta(1) with N->infinity, and finite N ~ 10^3-5000 captures the limit
    Table 1 fixes the order of limits; some experiments run P=1000 with N=1024 (Fig. 7), i.e., P/N = O(1); the paper appeals to O(1/sqrt(N)) fluctuations (App. C) rather than proving the corrections are negligible at these sizes.
  • domain assumption The aNTK fixed point determines the kernel predictor
    Abstract and Section 3.2 claim the fixed point equations define the kernel; App. F.2 states the fixed point only fixes the first two moments of p(h) and that the full marginal remains history dependent, so this assumption is only partially valid.
invented entities (3)
  • Adaptive Bayesian kernel PhiL (aNBK) independent evidence
    purpose: Data-dependent feature kernel defining the Bayesian adaptive predictor (Eqs. 4-5) as the solution of a min-max problem.
    Makes falsifiable predictions (test loss, layer-wise alignments, non-Gaussian preactivation densities) matched against finite-width Langevin simulations in Figs. 3, 4, 6, 7.
  • Adaptive tangent kernel K_aNTK independent evidence
    purpose: Fixed-point kernel defining the weight-decay gradient-flow predictor (Eqs. 10-11) through DMFT averages over h and gradient fields.
    Predicts test loss and preactivation/pre-gradient densities for finite-width GD-trained networks (Figs. 1, 5, 10-13).
  • Dual kernels hatPhi_l
    purpose: Conjugate matrix order parameters in the Bayesian action (Eqs. 6-8, 31) encoding feature-learning backreaction on layer densities.
    Internal saddle-point bookkeeping; not directly observable, although they tilt the single-site density Z_l (Eq. 7).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive kernel predictors from feature-learning infinite limits of neural networks." pith.science (2026). https://pith.science/paper/JCMSLQ5G

@misc{pith2026250207998,
  author       = {Pith},
  title        = {Pith review of: Adaptive kernel predictors from feature-learning infinite limits of neural networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JCMSLQ5G}},
  note         = {Machine review of arXiv:2502.07998}
}
read the original abstract

Previous influential work showed that infinite width limits of neural networks in the lazy training regime are described by kernel machines. Here, we show that neural networks trained in the rich, feature learning infinite-width regime in two different settings are also described by kernel machines, but with data-dependent kernels. For both cases, we provide explicit expressions for the kernel predictors and prescriptions to numerically calculate them. To derive the first predictor, we study the large-width limit of feature-learning Bayesian networks, showing how feature learning leads to task-relevant adaptation of layer kernels and preactivation densities. The saddle point equations governing this limit result in a min-max optimization problem that defines the kernel predictor. To derive the second predictor, we study gradient flow training of randomly initialized networks trained with weight decay in the infinite-width limit using dynamical mean field theory (DMFT). The fixed point equations of the arising DMFT defines the task-adapted internal representations and the kernel predictor. We compare our kernel predictors to kernels derived from lazy regime and demonstrate that our adaptive kernels achieve lower test loss on benchmark datasets.

Figures

Figures reproduced from arXiv: 2502.07998 by the authors.

Figure 1
Figure 1. Test network predictors of a two-layer MLP (width N = 5000) trained with P = 300 data of two-classes of CIFAR10 compared with the theoretical kernel regression predictors. Three panels are for different regimes of regularization λ and temperature 1/β. In all the three cases of λ, Kinit represent the network predictors at initialization, while Kfinal/Φfinal correspond either to aNTK at convergence for (a)/(b) or aNBK… view at source ↗
Figure 2
Figure 2. Linear networks with whitened data are determined by a set of kernel-label overlap matrices. (a) The overlap variables cℓ increase exponentially with ℓ with rate that depends on γ. Solid lines taken from Langevin dynamics on N = 1024 network. (b) The alignment of the final layer cL as a function of γ0 and L exhibits three distinct scaling regimes. (c) Examples of learned kernels (at each layer ℓ) in depth ℓ ∈ {L = 8… view at source ↗
Figure 3
Figure 3. Feature learning theories outperform lazy predictors for a two-layer MLP trained with Squared Loss (SL) on two classes of CIFAR10 (airplane vs automobile). (a) Test losses as a function of sample size P. Solid lines refer to theories, dashed lines to numerical simulations on a N = 5000 network. Blue is the NNGP lazy predictor; green is the deterministic NNGPK kernel predictor; orange is aNTK with feature learning st… view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Theory vs empirical kernels at each layer for a 5HL MLP with ϕ(h) ≡ tanh(h) learning P = 50 patterns of MNIST with y = {±1} P labels. Alignments A(Φℓ , Φℓ exp) = Tr(ΦℓΦℓ exp) ∥Φℓ∥∥Φℓ exp∥ between theory and empirical kernels are A(Φ1 , Φ1 exp) = 97%, A(Φ2 , Φ2 exp) = 8…
Figure 5
Figure 5. Figure 5: Test Loss as a function of sample size P for DMFT theories at convergence: two-layer MLP vs two-layer CNN trained on P animate/inanimate data on CIFAR10. Dashed lines are for theory, full-colored curves for empirical kernels. See Appendix E.1 and [PITH_FULL_IMAGE:figu…
Figure 6
Figure 6. Figure 6: (a) Bayesian theory (aNBK) and empirical (Langevin) adaptive kernels with feature learning strength γ0 and for P = 50 patterns of 0/1 classes of MNIST. C is the Gram matrix of data. (b) The preactivation distribution is in general non-Gaussian at each value of feature …
Figure 7
Figure 7. Figure 7: Bayesian L = 5 ReLU MLP trained on P = 1000 data of CIFAR10. Colored curve are experiments, dashed lines are the predictors calculated from Algorithm 1. (a) Kernel alignments A(Φℓ , yy⊤) at each layer vs feature strength. (b) Theory vs empirical kernels at each layer. …
Figure 8
Figure 8. Figure 8: (a) Bayesian two-layer MLP trained on a whitened covariance matrix Φ0 = I on P = 4 train points. Feature learning (γ0 > 0) leads to a non-Gaussian pre-activation distributions. Black-dashed curve is the lazy NNGP when h ∼ N (0, Φ0 ); red curve is the aNBK theory when h…
Figure 9
Figure 9. Figure 9: Finite-width effects in NTK and µP parameterizations. Blue curves are for different N values, showing that NNGP predictor is not consistent across network widths. We compare predictions of finite width N networks (as in (Pacelli et al., 2023)) to the NNGP infinite widt…
Figure 10
Figure 10. Figure 10: The training dynamics of two layer ReLU MLPs trained with weight decay. The richly trained networks achieve lower training and test errors at equal levels of regularization. Our theory can reproduce the final preactivation and pregradient densities in each setting. 10…
Figure 11
Figure 11. Figure 11: Training dynamics for a two layer CNN for varying richness γ0 on CIFAR-10 images. The images are turned into patches before computing cross-spatial correlations in the data. The training dynamics for infinite width networks (black) is compared to training finite width…
Figure 12
Figure 12. Figure 12: Pre-activation densities of a two-layer MLP trained with GD and weight decay at different times. Ligther colors represent the end of training. Dashed blue line is the theoretical prediction from the fixed point in the infinite time limit. F.2. Linear case In principle…
Figure 13
Figure 13. Figure 13: Pre-activation density of a two-layer MLP trained with P = 1 with a white covariance matrix. Dark colors correspond to early time training, being the weights initialized as Laplace distributed at t = 0. Light colors coincide with the end of training, when the system h…
Figure 14
Figure 14. Figure 14: Weight decay in the lazy training regime can cause a model to “unlearn” and reduce its output after the features start to decay. Provided γ0 is sufficiently large compared to λ, however, the final predictor will still be nontrivial, unlike the zero predictor obtained …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Statistical physics of deep learning: Optimal learning of a multi-layer perceptron near interpolation

    stat.ML 2025-10 conditional novelty 8.0 of 10

    A replica/HCIZ theory predicts the Bayes-optimal generalization error of proportional-width MLPs near interpolation and discovers layer-wise specialization transitions that make deeper targets harder to learn.

  2. Width-Robust Learnability in Mean-Field Bayesian Neural Networks

    stat.ML 2026-07 conditional novelty 7.0 of 10

    For fixed-depth mean-field Bayesian nets on the Boolean cube, poly-sample learnability at infinite width equals poly-width learnability equals poly-bounded reduced entropy.

  3. Dynamics of neural scaling laws in random feature regression with powerlaw-distributed kernel eigenvalues

    cond-mat.dis-nn 2026-02 conditional novelty 6.0 of 10

    A single dynamical mean-field theory unifies Bayesian, gradient-flow, and Langevin training of random-feature regression and explains finite-time generalization error on power-law spectra.

Reference graph

Works this paper leans on

65 extracted references · 28 canonical work pages · cited by 3 Pith papers

  1. [1]

    Advani, M. S. and Saxe, A. M. High-dimensional dynamics of generalization error in neural networks, 2017. URL https://arxiv.org/abs/1710.03667

  2. [2]

    Why bigger is not always better: on finite and infinite neural networks

    Aitchison, L. Why bigger is not always better: on finite and infinite neural networks, 2020. URL https://arxiv.org/abs/1910.08013

  3. [3]

    Local kernel renormalization as a mechanism for feature learning in overparametrized convolutional neural networks

    Aiudi, R., Pacelli, R., Vezzani, A., Burioni, R., and Rotondo, P. Local kernel renormalization as a mechanism for feature learning in overparametrized convolutional neural networks. Nature Communications, 16, 2023. URL https://api.semanticscholar.org/CorpusID:260125263

  4. [4]

    A convergence analysis of gradient descent for deep linear neural networks, 2019 a

    Arora, S., Cohen, N., Golowich, N., and Hu, W. A convergence analysis of gradient descent for deep linear neural networks, 2019 a . URL https://arxiv.org/abs/1810.02281

  5. [5]

    S., Hu, W., Li, Z., Salakhutdinov, R., and Wang, R

    Arora, S., Du, S. S., Hu, W., Li, Z., Salakhutdinov, R., and Wang, R. On exact computation with an infinitely wide neural net, 2019 b . URL https://arxiv.org/abs/1904.11955

  6. [6]

    Cugliandolo-Kurchan equations for dynamics of Spin-Glasses

    Arous, G. B., Dembo, A., and Guionnet, A. Cugliandolo-kurchan equations for dynamics of spin-glasses, 2004. URL https://arxiv.org/abs/math/0409273

  7. [7]

    Neural networks as kernel learners: The silent alignment effect, 2021

    Atanasov, A., Bordelon, B., and Pehlevan, C. Neural networks as kernel learners: The silent alignment effect, 2021. URL https://arxiv.org/abs/2111.00034

  8. [8]

    Predictive power of a bayesian effective action for fully connected one hidden layer neural networks in the proportional limit

    Baglioni, P., Pacelli, R., Aiudi, R., Di Renzo, F., Vezzani, A., Burioni, R., and Rotondo, P. Predictive power of a bayesian effective action for fully connected one hidden layer neural networks in the proportional limit. Phys. Rev. Lett., 133: 0 027301, Jul 2024. doi:10.1103/PhysRevLett.133.027301. URL https://link.aps.org/doi/10.1103/PhysRevLett.133.027301

Show all 65 references
  1. [9]

    Feature learning in finite-width bayesian deep linear networks with multiple outputs and convolutional layers, 2024

    Bassetti, F., Gherardi, M., Ingrosso, A., Pastore, M., and Rotondo, P. Feature learning in finite-width bayesian deep linear networks with multiple outputs and convolutional layers, 2024. URL https://arxiv.org/abs/2406.03260

  2. [10]

    and Pehlevan, C

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

  3. [11]

    and Pehlevan, C

    Bordelon, B. and Pehlevan, C. Dynamics of finite width kernel and prediction fluctuations in mean field neural networks. Advances in Neural Information Processing Systems, 36, 2024

  4. [12]

    How feature learning can improve neural scaling laws, 2024

    Bordelon, B., Atanasov, A., and Pehlevan, C. How feature learning can improve neural scaling laws, 2024. URL https://arxiv.org/abs/2409.17858

  5. [13]

    J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., Vander P las, J., Wanderman- M ilne, S., and Zhang, Q

    Bradbury, J., Frostig, R., Hawkins, P., Johnson, M. J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., Vander P las, J., Wanderman- M ilne, S., and Zhang, Q. JAX : composable transformations of P ython+ N um P y programs, 2018. URL http://github.com/jax-ml/jax

  6. [14]

    On lazy training in differentiable programming, 2020

    Chizat, L., Oyallon, E., and Bach, F. On lazy training in differentiable programming, 2020. URL https://arxiv.org/abs/1812.07956

  7. [15]

    and Saul, L

    Cho, Y. and Saul, L. Kernel methods for deep learning. In Bengio, Y., Schuurmans, D., Lafferty, J., Williams, C., and Culotta, A. (eds.), Advances in Neural Information Processing Systems, volume 22. Curran Associates, Inc., 2009 a . URL https://proceedings.neurips.cc/paper_fi...

  8. [16]

    and Saul, L

    Cho, Y. and Saul, L. Kernel methods for deep learning. Advances in neural information processing systems, 22, 2009 b

  9. [17]

    Bayes-optimal learning of deep random networks of extensive-width

    Cui, H., Krzakala, F., and Zdeborov \'a , L. Bayes-optimal learning of deep random networks of extensive-width. In International Conference on Machine Learning, pp.\ 6468--6521. PMLR, 2023

  10. [18]

    Dynamics as a substitute for replicas in systems with quenched random impurities

    De Dominicis, C. Dynamics as a substitute for replicas in systems with quenched random impurities. Phys. Rev. B, 18: 0 4913--4919, Nov 1978. doi:10.1103/PhysRevB.18.4913. URL https://link.aps.org/doi/10.1103/PhysRevB.18.4913

  11. [19]

    Matthews, A

    de G. Matthews, A. G., Rowland, M., Hron, J., Turner, R. E., and Ghahramani, Z. Gaussian process behaviour in wide deep neural networks, 2018. URL https://arxiv.org/abs/1804.11271

  12. [20]

    Every model learned by gradient descent is approximately a kernel machine

    Domingos, P. Every model learned by gradient descent is approximately a kernel machine. arXiv preprint arXiv:2012.00152, 2020

  13. [21]

    Critical feature learning in deep neural networks, 2024

    Fischer, K., Lindner, J., Dahmen, D., Ringel, Z., Krämer, M., and Helias, M. Critical feature learning in deep neural networks, 2024. URL https://arxiv.org/abs/2405.10761

  14. [22]

    Disentangling feature and lazy training in deep neural networks

    Geiger, M., Spigler, S., Jacot, A., and Wyart, M. Disentangling feature and lazy training in deep neural networks. Journal of Statistical Mechanics: Theory and Experiment, 2020 0 (11): 0 113301, November 2020. ISSN 1742-5468. doi:10.1088/1742-5468/abc4de. URL http://dx.doi.org...

  15. [23]

    and Zlokapa, A

    Hanin, B. and Zlokapa, A. Bayesian interpolation with deep linear networks. Proceedings of the National Academy of Sciences, 120 0 (23): 0 e2301345120, 2023

  16. [24]

    Hestness, J., Narang, S., Ardalani, N., Diamos, G., Jun, H., Kianinejad, H., Patwary, M. M. A., Yang, Y., and Zhou, Y. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017

  17. [25]

    Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., Casas, D. d. L., Hendricks, L. A., Welbl, J., Clark, A., et al. Training compute-optimal large language models. arXiv preprint arXiv:2203.15556, 2022

  18. [26]

    Statistical mechanics of transfer learning in fully connected networks in the proportional limit

    Ingrosso, A., Pacelli, R., Rotondo, P., and Gerace, F. Statistical mechanics of transfer learning in fully connected networks in the proportional limit. Physical Review Letters, 134 0 (17): 0 177301, 2025

  19. [27]

    Neural tangent kernel: Convergence and generalization in neural networks, 2020

    Jacot, A., Gabriel, F., and Hongler, C. Neural tangent kernel: Convergence and generalization in neural networks, 2020. URL https://arxiv.org/abs/1806.07572

  20. [28]

    Saddle-to-saddle dynamics in deep linear networks: Small initialization training, symmetry, and sparsity, 2022

    Jacot, A., Ged, F., Şimşek, B., Hongler, C., and Gabriel, F. Saddle-to-saddle dynamics in deep linear networks: Small initialization training, symmetry, and sparsity, 2022. URL https://arxiv.org/abs/2106.15933

  21. [29]

    B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D

    Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020

  22. [30]

    Statistical physics of particles

    Kardar, M. Statistical physics of particles. Cambridge University Press, 2007

  23. [31]

    S., Pennington, J., and Sohl-Dickstein, J

    Lee, J., Bahri, Y., Novak, R., Schoenholz, S. S., Pennington, J., and Sohl-Dickstein, J. Deep neural networks as gaussian processes, 2018. URL https://arxiv.org/abs/1711.00165

  24. [32]

    S., Bahri, Y., Novak, R., Sohl-Dickstein, J., and Pennington, J

    Lee, J., Xiao, L., Schoenholz, S. S., Bahri, Y., Novak, R., Sohl-Dickstein, J., and Pennington, J. Wide neural networks of any depth evolve as linear models under gradient descent *. Journal of Statistical Mechanics: Theory and Experiment, 2020 0 (12): 0 124002, December 2020....

  25. [33]

    and Gur-Ari, G

    Lewkowycz, A. and Gur-Ari, G. On the training dynamics of deep networks with l\_2 regularization. Advances in Neural Information Processing Systems, 33: 0 4790--4799, 2020

  26. [34]

    and Sompolinsky, H

    Li, Q. and Sompolinsky, H. Statistical mechanics of deep linear neural networks: The backpropagating kernel renormalization. Physical Review X, 11 0 (3), September 2021. ISSN 2160-3308. doi:10.1103/physrevx.11.031059. URL http://dx.doi.org/10.1103/PhysRevX.11.031059

  27. [35]

    and Sompolinsky, H

    Li, Q. and Sompolinsky, H. Globally gated deep linear networks. Advances in Neural Information Processing Systems, 35: 0 34789--34801, 2022

  28. [36]

    Emergence in non-neural models: grokking modular arithmetic via average gradient outer product

    Mallinar, N., Beaglehole, D., Zhu, L., Radhakrishnan, A., Pandit, P., and Belkin, M. Emergence in non-neural models: grokking modular arithmetic via average gradient outer product. arXiv preprint arXiv:2407.20199, 2024

  29. [37]

    C., Siggia, E

    Martin, P. C., Siggia, E. D., and Rose, H. A. Statistical dynamics of classical systems. Phys. Rev. A, 8: 0 423--437, Jul 1973. doi:10.1103/PhysRevA.8.423. URL https://link.aps.org/doi/10.1103/PhysRevA.8.423

  30. [38]

    A mean field view of the landscape of two-layer neural networks

    Mei, S., Montanari, A., and Nguyen, P.-M. A mean field view of the landscape of two-layer neural networks. Proceedings of the National Academy of Sciences, 115 0 (33): 0 E7665--E7671, 2018. doi:10.1073/pnas.1806579115. URL https://www.pnas.org/doi/abs/10.1073/pnas.1806579115

  31. [39]

    Mingard, C., Valle-Pérez, G., Skalse, J., and Louis, A. A. Is sgd a bayesian sampler? well, almost, 2020. URL https://arxiv.org/abs/2006.15191

  32. [40]

    and Ringel, Z

    Naveh, G. and Ringel, Z. A self consistent theory of gaussian processes captures feature learning effects in finite cnns. Advances in Neural Information Processing Systems, 34: 0 21352--21364, 2021

  33. [41]

    Predicting the outputs of finite deep neural networks trained with noisy gradients

    Naveh, G., Ben David, O., Sompolinsky, H., and Ringel, Z. Predicting the outputs of finite deep neural networks trained with noisy gradients. Physical Review E, 104 0 (6), December 2021. ISSN 2470-0053. doi:10.1103/physreve.104.064301. URL http://dx.doi.org/10.1103/PhysRevE.104.064301

  34. [42]

    Neal, R. M. Bayesian learning for neural networks. PhD thesis, CAN, 1995. AAINN02676

  35. [43]

    A., Pennington, J., and Sohl-Dickstein, J

    Novak, R., Bahri, Y., Abolafia, D. A., Pennington, J., and Sohl-Dickstein, J. Sensitivity and generalization in neural networks: an empirical study, 2018. URL https://arxiv.org/abs/1802.08760

  36. [44]

    A statistical mechanics framework for bayesian deep neural networks beyond the infinite-width limit

    Pacelli, R., Ariosto, S., Pastore, M., Ginelli, F., Gherardi, M., and Rotondo, P. A statistical mechanics framework for bayesian deep neural networks beyond the infinite-width limit. Nature Machine Intelligence, 5 0 (12): 0 1497–1507, December 2023. ISSN 2522-5839. doi:10.1038...

  37. [45]

    Mechanism of feature learning in deep fully connected networks and kernel machines that recursively learn features

    Radhakrishnan, A., Beaglehole, D., Pandit, P., and Belkin, M. Mechanism of feature learning in deep fully connected networks and kernel machines that recursively learn features. arXiv preprint arXiv:2212.13881, 2022

  38. [46]

    Rasmussen, C. E. and Williams, C. K. I. Gaussian Processes for Machine Learning. The MIT Press, 2006

  39. [47]

    A., Yaida, S., and Hanin, B

    Roberts, D. A., Yaida, S., and Hanin, B. The principles of deep learning theory, volume 46. Cambridge University Press Cambridge, MA, USA, 2022

  40. [48]

    and Vanden-Eijnden, E

    Rotskoff, G. and Vanden-Eijnden, E. Trainability and accuracy of artificial neural networks: An interacting particle system approach. Communications on Pure and Applied Mathematics, 75 0 (9): 0 1889--1935, 2022

  41. [49]

    A unified approach to feature learning in bayesian neural networks

    Rubin, N., Ringel, Z., Seroussi, I., and Helias, M. A unified approach to feature learning in bayesian neural networks. In High-dimensional Learning Dynamics 2024: The Emergence of Structure and Reasoning, 2024 a . URL https://openreview.net/forum?id=ZmOSJ2MV2R

  42. [50]

    Grokking as a first order phase transition in two layer networks

    Rubin, N., Seroussi, I., and Ringel, Z. Grokking as a first order phase transition in two layer networks. In The Twelfth International Conference on Learning Representations, 2024 b . URL https://openreview.net/forum?id=3ROGsTX3IR

  43. [51]

    M., McClelland, J

    Saxe, A. M., McClelland, J. L., and Ganguli, S. Exact solutions to the nonlinear dynamics of learning in deep linear neural networks, 2014. URL https://arxiv.org/abs/1312.6120

  44. [52]

    and Smola, A

    Scholkopf, B. and Smola, A. J. Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond. MIT Press, Cambridge, MA, USA, 2001. ISBN 0262194759

  45. [53]

    Separation of scales and a thermodynamic description of feature learning in some cnns

    Seroussi, I., Naveh, G., and Ringel, Z. Separation of scales and a thermodynamic description of feature learning in some cnns. Nature Communications, 14 0 (1): 0 908, 2023

  46. [54]

    Order parameters and phase transitions of continual learning in deep neural networks

    Shan, H., Li, Q., and Sompolinsky, H. Order parameters and phase transitions of continual learning in deep neural networks. arXiv preprint arXiv:2407.10315, 2024

  47. [55]

    and Zippelius, A

    Sompolinsky, H. and Zippelius, A. Dynamic theory of the spin-glass phase. Phys. Rev. Lett., 47: 0 359--362, Aug 1981. doi:10.1103/PhysRevLett.47.359. URL https://link.aps.org/doi/10.1103/PhysRevLett.47.359

  48. [56]

    and Sompolinsky, H

    van Meegen, A. and Sompolinsky, H. Coding schemes in neural networks learning classification tasks, 2024. URL https://arxiv.org/abs/2406.16689

  49. [57]

    Limitations of the ntk for understanding generalization in deep learning

    Vyas, N., Bansal, Y., and Nakkiran, P. Limitations of the ntk for understanding generalization in deep learning. arXiv preprint arXiv:2206.10012, 2022

  50. [58]

    Feature-learning networks are consistent across widths at realistic scales, 2023

    Vyas, N., Atanasov, A., Bordelon, B., Morwani, D., Sainathan, S., and Pehlevan, C. Feature-learning networks are consistent across widths at realistic scales, 2023. URL https://arxiv.org/abs/2305.18411

  51. [59]

    More than a toy: Random matrix models predict how real-world neural representations generalize, 2022

    Wei, A., Hu, W., and Steinhardt, J. More than a toy: Random matrix models predict how real-world neural representations generalize, 2022. URL https://arxiv.org/abs/2203.06176

  52. [60]

    and Teh, Y

    Welling, M. and Teh, Y. W. Bayesian learning via stochastic gradient langevin dynamics. In Proceedings of the 28th International Conference on International Conference on Machine Learning, ICML'11, pp.\ 681–688, Madison, WI, USA, 2011. Omnipress. ISBN 9781450306195

  53. [61]

    X., Robeyns, M., Milsom, E., Anson, B., Schoots, N., and Aitchison, L

    Yang, A. X., Robeyns, M., Milsom, E., Anson, B., Schoots, N., and Aitchison, L. A theory of representation learning gives a deep generalisation of kernel methods. In International Conference on Machine Learning, pp.\ 39380--39415. PMLR, 2023

  54. [62]

    and Hu, E

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

  55. [63]

    J., Babuschkin, I., Sidor, S., Liu, X., Farhi, D., Ryder, N., Pachocki, J., Chen, W., and Gao, J

    Yang, G., Hu, E. J., Babuschkin, I., Sidor, S., Liu, X., Farhi, D., Ryder, N., Pachocki, J., Chen, W., and Gao, J. Tensor programs v: Tuning large neural networks via zero-shot hyperparameter transfer, 2022. URL https://arxiv.org/abs/2203.03466

  56. [64]

    A., Canatar, A., Ruben, B

    Zavatone-Veth, J. A., Canatar, A., Ruben, B. S., and Pehlevan, C. Asymptotics of representation learning in finite bayesian neural networks*. Journal of Statistical Mechanics: Theory and Experiment, 2022 0 (11): 0 114008, November 2022 a . ISSN 1742-5468. doi:10.1088/1742-5468...

  57. [65]

    A., Tong, W

    Zavatone-Veth, J. A., Tong, W. L., and Pehlevan, C. Contrasting random and learned features in deep bayesian linear regression. Physical Review E, 105 0 (6): 0 064118, 2022 b

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.