Pith. sign in

REVIEW 3 major objections 6 minor 17 references

Reactivation: Empirical NTK Dynamics Under Task Shifts

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Task shifts force the neural tangent kernel of past data to move abruptly, even in wide, supposedly lazy networks, with the effect size set by how semantically new the incoming task is.

desk verdict A genuinely new empirical observation about NTK dynamics at task boundaries, but the 'large scale' claim needs a scaling fit and the estimator needs seeds and error bars. read the letter →

arxiv 2507.16039 v2 pith:73LLGYKX submitted 2025-07-21 cs.LG cs.AI

classification cs.LGcs.AI
keywords neuraltangentkernelcontinuallearninglazyregimefeaturealignmenttasksimilaritydistributionshiftvelocity
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

This paper tries to show that the neural tangent kernel (NTK)—the object that supposedly stays frozen when a network is wide enough to be in the 'lazy' regime—does not stay frozen in continual learning. At the moment a model switches from one task to the next, the kernel of the first task's data undergoes an abrupt, temporary reorganization: a drop-and-recovery 'check-mark' in its top eigenvalue, a spike in kernel velocity, and a shift in kernel alignment. The phenomenon appears across widths from 64 to 2048, across learning rates from 1e-3 to 1e-5, and in both feature-learning and lazy parametrizations. Its size tracks how semantically new the second task is: introducing genuinely new classes causes large kernel changes, while reweighting known classes causes almost none. If this is right, static-kernel approximations in continual learning theory are missing the most important moment of training.

What carries the argument

The machinery is the empirical NTK of the first-task data, computed from gradient inner products $\Theta_t(x_i, x_j) = \nabla_\theta f_t(x_i)^\top \nabla_\theta f_t(x_j)$ on a 32-sample batch, plus four derived diagnostics: its maximum eigenvalue (spectral norm), kernel distance $S(\Theta, \Theta') = 1 - \mathrm{CKA}(\Theta, \Theta')$, kernel velocity $v(t) = S(\Theta_t, \Theta_{t+dt})/dt$, and kernel alignment $A(t) = \mathrm{CKA}(\Theta_t, yy^\top)$. The eigenmode recursion $e_{t+1} = (1 - \eta\Lambda)e_t$ links the spectral norm to convergence speed. The paper's argument works by watching these diagnostics on old-task data across a task boundary, comparing widths, learning rates, and two kinds of distribution shift.

What would settle it

Recompute the NTK diagnostics with the same fixed 32 images at every checkpoint, with freshly sampled 32-image batches, and with 512 or more images; if the drop-recovery check-mark and velocity spike disappear in the fixed-batch or large-batch runs, the reactivation is an artifact of the measurement window rather than a property of training.

Watch

Extended reading notes

Core claim

The central claim is that task boundaries reactivate feature learning even where single-task theory says it should be dormant. Concretely, with all metrics evaluated on the first task's data, the NTK's maximum eigenvalue drops sharply when the second task begins and then recovers, the kernel distance from initialization jumps, and kernel velocity—defined as $S(\Theta_t, \Theta_{t+dt})/dt$ with $S$ a CKA-based distance—spikes and decays back to its pre-switch level within a few epochs. The same pattern holds under NTK parametrization with learning rate scaled as $1/N$, where the kernel is nearly static during the first task itself, so the authors take the switch-triggered motion as evidence of reactivation rather than ordinary training drift. They further claim a monotone, sublinear relation between the number of new classes in the second task and the size of the kernel disruption, and contrast this with frequency-only shifts, where no reactivation occurs.

Load-bearing premise

The whole phenomenon is read off an empirical NTK estimated from 32 random images of the first task, and the claim stands on that small-sample kernel being a stable, unbiased stand-in for the true kernel of that task.

Editorial extensions

If this is right

  • At the widths and learning rates tested (64 to 2048 channels, learning rates $10^{-3}$ to $10^{-5}$), a static NTK is not a valid approximation across a task boundary, so lazy-regime guarantees for continual learning do not transfer directly.
  • The magnitude of reactivation is predictable from semantic overlap: introducing new classes moves the kernel far more than reweighting known ones, so representation drift at a boundary is partially forecastable.
  • Widening the network shrinks but does not remove the reactivation; the drop-recovery shape persists at width 2048 and on ImageNet100, so scale alone does not restore kernel stationarity in non-stationary settings.
  • The transient nature of the spike, with velocity decaying back to baseline within a few epochs, suggests the network passes through a short adaptation window at each boundary, after which dynamics settle again.
  • Frequency-only distribution shifts produce smooth, minor kernel change, so continual-learning theories must distinguish 'new content' from 'reweighted content' rather than treating all shifts alike.

Reading between the lines

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

  • A testable extension the paper leaves implicit: if reactivation is real, the same spike should appear when measuring the NTK on the new task's data from the first iteration of task 2, and its magnitude should predict how much accuracy on task 1 drops afterward, connecting kernel velocity at the boundary to forgetting.
  • The check-mark shape resembles the transient sharpening seen in loss-landscape studies through the NTK-Hessian link, so connecting reactivation to edge-of-stability dynamics is a natural next step not taken here.
  • The 32-sample estimator needs a stability check before the phenomenon is treated as quantitative: re-measuring with fixed versus freshly sampled batches, and across output logits, would settle whether the spike is intrinsic or an estimator artifact.
  • If semantic novelty is the driver, then task-construction choices in continual learning benchmarks, such as how many new classes appear per task, directly control how much feature re-learning occurs, so benchmark design could deliberately tune reactivation.
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 / 6 minor

Summary. The paper presents an empirical study of Neural Tangent Kernel (NTK) dynamics in continual learning. On CIFAR-10 and ImageNet100, a CNN is trained on a sequence of tasks while the paper measures, on the first-task data only, the NTK max eigenvalue, kernel distance 1−CKA, kernel velocity, and kernel alignment. The central observation is that at the moment of a task switch the NTK of the past data changes abruptly — a velocity spike and a drop-recovery ('check-mark') in the max eigenvalue — across widths 64–2048, learning rates 1e-3–1e-5, and in both feature-learning and lazy (learning rate scaled by 1/N) regimes. The paper also shows that introducing new classes drives larger NTK changes than reweighting existing classes, and that the magnitude of the reactivation decays with width. The authors argue these findings challenge static-kernel approximations in continual learning.

Significance. If the central observation holds, the paper identifies a real and underappreciated phenomenon: task boundaries transiently move finite-width networks away from lazy dynamics, so static-kernel theories of continual learning miss a boundary effect. The paper's strengths are its direct, operational definition of reactivation through kernel metrics on fixed first-task data, the consistency of the pattern across many widths, learning rates, and datasets, and the clean comparison between semantic novelty and frequency shift, which yields a well-ordered effect. These make the phenomenon falsifiable. However, the significance of the large-scale claim is currently unquantified: the paper's own scaling figures show a decreasing amplitude, and without a fitted intercept the asymptotic relevance is not established. The underspecified NTK estimator also needs to be pinned down before the quantitative conclusions can be trusted.

major comments (3)
  1. [Appendix C.4, Fig. 12; Sec. 2.2] The claim that reactivation persists 'even at large scale' is a central conclusion, but the only scaling evidence is a monotonic decrease of the response amplitude with width from 64 to 2048 (Fig. 12; Appendix C.4), with a similar smoothing in the ImageNet curves (Fig. 11). Without a fitted asymptote, the data are equally consistent with a finite-width effect that vanishes in the infinite-width limit. I request a scaling analysis: define the reactivation amplitude (e.g., kernel-velocity peak or max-eigenvalue drop at the switch), average over seeds, and fit Δ(N) = a + b N^{-α} with a confidence interval on the intercept a. If a is consistent with zero, the headline claim should be weakened to a finite-width phenomenon.
  2. [Appendix A.3] The NTK estimator is underspecified in two ways. First, for a multi-class network the NTK is a tensor indexed by output logits; the paper does not state which scalar output or loss gradient is recorded when a 'single sample' is fed. If the recorded vector is the cross-entropy loss gradient rather than the Jacobian of a logit, the computed matrix is not the NTK. Second, it is not stated whether the same 32 random samples are reused at every time step; if the batch is re-sampled, the velocity S(Θ_t, Θ_{t+dt}) includes sampling noise that could inflate the task-switch spike. Please specify the exact computation and report a fixed-vs-resampled batch comparison.
  3. [Sec. 2.2, Figs. 1–4, 7–12] No seed variance or error bars are reported for any of the central curves, although Sec. 2 says experiments were run 'across multiple seeds.' The qualitative check-mark and velocity spike may be robust in the shown single runs, but the decreasing amplitude with width makes run-to-run variation a first-order concern. Please report mean ± standard deviation (or individual seeds) for at least the reactivation amplitude and for the key curves in a few representative settings.
minor comments (6)
  1. [Sec. 2.3] In Experiment 1, 'varying i ∈ [0, 1]' should be 'i ∈ {0, …, 10}' (or [0, 10]) to match the experiments E0–E10 and the similarity formula.
  2. [Appendix A.2] The displayed CKA formula is uncentered linear alignment; if centering is intended, the definition should use centered kernel matrices, and the text should say so.
  3. [Appendix C.3, Fig. 11] The caption for the second row lists 20, 50, and 100 epochs per task, but panels (d)–(f) are labeled 10, 20, and 100 epochs; please make the numbers consistent.
  4. [Sec. 2.1, Fig. 7] Please define 'step' and 'dt' in the main text: one step corresponds to 10 training iterations, and the velocity is computed with dt = 10 steps (100 iterations) according to Fig. 7; stating this only in captions is insufficient for reproducibility.
  5. [References] References [8] and [9] appear to be the same paper (Jacot, Gabriel, Hongler, NeurIPS 2018); please consolidate to a single entry.
  6. [Appendix B.1] The description 'Pytorch standard parametrization' and 'very similar to NTK parametrization with only factor-wise difference' is too vague to reproduce; please specify the exact initialization, the learning-rate scaling, and any output scaling for each regime.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reactivation result is an empirical measurement of NTK metrics, not a consequence fitted from or defined by the conclusion.

full rationale

The paper is purely empirical: every central claim is a direct measurement of NTK-derived quantities on first-task data. Section 2.1 defines kernel spectral norm, kernel distance, kernel velocity, and kernel alignment without reference to the phenomenon being claimed; Section 2.2 reports the observed spike in kernel velocity and the check-mark trajectory in NTK max eigenvalue and explicitly states, "We refer to this phenomenon as the re-activation of feature learning." This is a naming of a measured pattern, not a derivation from the conclusion. The theoretical appendix equations (3)-(4) are standard per-eigenmode error dynamics and are used only to motivate why the NTK spectrum matters; they do not encode the task-switch result. There are no fitted parameters renamed as predictions, no uniqueness theorem imported from the authors' prior work, and no load-bearing self-citation chain. The cited references (Fort et al., Zhou et al., Yang, Yang and Hu) provide external context for NTK evolution and parametrization, but the observed reactivation at task boundaries is not forced by any of those citations. The skeptical concerns—that Fig. 12 and Appendix C.4 show reactivation amplitude decreasing with width, and that Appendix A.3 uses 32 random samples with no stated seed or same-batch protocol—are legitimate evidence-quality and extrapolation concerns, not circularity: they question whether the finite-width observation persists at the large-width limit or whether the measurement is stable, but they do not show that any claim reduces by construction to its inputs. Under the stated criteria, a non-finding with score 0 is appropriate.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper makes no fitted-model claims; its conclusions rest on direct NTK measurements. The main load-bearing assumptions are the fidelity of the 32-sample NTK estimate, the validity of the Kaiming-plus-1/N-LR approximation to the lazy regime, and the transfer of squared-loss spectral analysis to cross-entropy training.

assumptions (4)
  • standard math The standard NTK framework (Jacot et al.) relating gradient descent to kernel evolution holds for the finite-width CNN studied.
    Invoked in Sec 2.1 and Appendix A to justify using NTK metrics as descriptors of training dynamics.
  • domain assumption Kaiming Normal initialization with learning rate scaled as 1/N approximates the NTK parametrization and induces the lazy regime at the tested widths.
    Appendix B.1 states this approximation; if false, the claim that observed changes occur 'under a lazy regime' is weakened.
  • domain assumption A 32-sample random batch yields a reliable estimate of the NTK trajectory.
    Appendix A.3 asserts this based on a consistency check across sample sizes 10/20/100, but does not establish that the estimate is unbiased or that the batch is fixed across time.
  • domain assumption The squared-loss eigenmode analysis (Eqs. 3-4) transfers to cross-entropy training, so that NTK spectral norm controls convergence speed in the experiments.
    Appendix A.1 derives error decay only for squared loss, while all experiments use cross-entropy loss (Appendix B.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reactivation: Empirical NTK Dynamics Under Task Shifts." pith.science (2026). https://pith.science/paper/73LLGYKX

@misc{pith2026250716039,
  author       = {Pith},
  title        = {Pith review of: Reactivation: Empirical NTK Dynamics Under Task Shifts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/73LLGYKX}},
  note         = {Machine review of arXiv:2507.16039}
}
read the original abstract

The Neural Tangent Kernel (NTK) offers a powerful tool to study the functional dynamics of neural networks. In the so-called lazy, or kernel regime, the NTK remains static during training and the network function is linear in the static neural tangents feature space. The evolution of the NTK during training is necessary for feature learning, a key driver of deep learning success. The study of the NTK dynamics has led to several critical discoveries in recent years, in generalization and scaling behaviours. However, this body of work has been limited to the single task setting, where the data distribution is assumed constant over time. In this work, we present a comprehensive empirical analysis of NTK dynamics in continual learning, where the data distribution shifts over time. Our findings highlight continual learning as a rich and underutilized testbed for probing the dynamics of neural training. At the same time, they challenge the validity of static-kernel approximations in theoretical treatments of continual learning, even at large scale.

Figures

Figures reproduced from arXiv: 2507.16039 by the authors.

Figure 1
Figure 1. Comparison of NTK Max Eigenvalue, Kernel Distance from initialization, Kernel Align￾ment and Kernel velocity across different widths with (Row 1) and without (Row 2) feature learning. The measurements are done on the first task before and after the task switch, with 5 classes in each task. One step corresponds to 10 iterations during training. 0 5000 10000 15000 20000 Step 0 2 4 6 NTK Max Eigenvalue 1e6 LR 1e-03 1e-… view at source ↗
Figure 2
Figure 2. Comparison of NTK Max Eigenvalue, Kernel Distance from initialization, Kernel Align￾ment and Kernel velocity across different learning rates at fixed width (N = 2048). We find that the behavior at task switch critically depends on the learning rate. Notably, the checkmark shape is more visible for higher learning rates, while for lower learning rates, it is col￾lapsed on the first few steps after the task switch. Mo… view at source ↗
Figure 3
Figure 3. NTK Dynamics of the first task with concept-based distribution shift. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: NTK Dynamics of the first task with frequency-based distribution shift. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: NTK dynamics computed with different sample sizes ( [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Kernel distance in Experiment 1 and a zoomed-in view around the task switch. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Comparison of different metrics across network widths for CNN trained on CIFAR10 with learning rate 0.001. The number of epochs per task is set to 160. (a) Test accuracy, (b) Alignment, (c) Kernel distance, (d) Maximum eigenvalue of NTK, (e) Kernel velocity with dt=10,…
Figure 8
Figure 8. Figure 8: Comparison of different metrics across network widths for CNN trained on CIFAR10 with learning rate 0.0001. The number of epochs per task is set to 160. (a) Test accuracy, (b) Alignment, (c) Kernel distance, (d) Maximum eigenvalue of NTK, (e) Kernel velocity with dt=10…
Figure 9
Figure 9. Figure 9: Comparison of different metrics across network widths for CNN trained on CIFAR10 with learning rate 0.00001. The number of epochs per task is set to 160. (a) Test accuracy, (b) Alignment, (c) Kernel distance, (d) Maximum eigenvalue of NTK, (e) Kernel velocity with dt=1…
Figure 10
Figure 10. Figure 10: Comparison of different metrics across network widths for CNN trained on CIFAR10 during multiple task switches. The number of epochs per task is set to 160. (a) Test accuracy, (b) Alignment, (c) Kernel distance, (d) Maximum eigenvalue of NTK, (e) Kernel velocity with …
Figure 11
Figure 11. Figure 11: The effect of width and number of epochs per task on NTK spectrum on ImageNet100. The first row, left to right width 250, 500 and 1000 respectively, the second row, left to right epoch number per task 20, 50 and 100 respectively. 0 5000 10000 15000 20000 Step 0.0 0.2 …
Figure 12
Figure 12. Figure 12: Comparison of different metrics across network widths for CNN trained on CIFAR10 with NTK parametrization. The number of epochs per task is set to 160. (a) Test accu￾racy, (b) Alignment, (c) Kernel distance, (d) Maximum eigenvalue of NTK, (e) Kernel velocity with dt=1…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 12 canonical work pages

  1. [1]

    Implicit regularization via neural feature alignment

    Aristide Baratin, Thomas George, C ´esar Laurent, R Devon Hjelm, Guillaume Lajoie, Pascal Vincent, and Simon Lacoste-Julien. Implicit regularization via neural feature alignment. In International Conference on Artificial Intelligence and Statistics , pages 2269–2277. PMLR, 2021

  2. [2]

    Generalisation guarantees for continual learning with orthogonal gradient descent

    Mehdi Abbana Bennani, Thang Doan, and Masashi Sugiyama. Generalisation guarantees for continual learning with orthogonal gradient descent. arXiv preprint arXiv:2006.11942, 2020

  3. [3]

    Gradient de- scent on neural networks typically occurs at the edge of stability

    Jeremy Cohen, Simran Kaur, Yuanzhi Li, J Zico Kolter, and Ameet Talwalkar. Gradient de- scent on neural networks typically occurs at the edge of stability. In International Confer- ence on Learning Representations , 2021. URL https://openreview.net/forum? id=jh-rTtvkGeM

  4. [4]

    Algorithms for learning kernels based on centered alignment

    Corinna Cortes, Mehryar Mohri, and Afshin Rostamizadeh. Algorithms for learning kernels based on centered alignment. The Journal of Machine Learning Research , 13(1):795–828, 2012

  5. [5]

    A Theoretical Analysis of Catastrophic Forgetting through the NTK Overlap Matrix

    Thang Doan, Mehdi Bennani, Bogdan Mazoure, Guillaume Rabusseau, and Pierre Alquier. A theoretical analysis of catastrophic forgetting through the ntk overlap matrix, 2021. URL https://arxiv.org/abs/2010.04003

  6. [6]

    Deep learning versus kernel learning: an empirical study of loss landscape geometry and the time evolution of the neural tangent kernel

    Stanislav Fort, Gintare Karolina Dziugaite, Mansheej Paul, Sepideh Kharaghani, Daniel M Roy, and Surya Ganguli. Deep learning versus kernel learning: an empirical study of loss landscape geometry and the time evolution of the neural tangent kernel. Advances in Neural Information Processing Systems, 33:5850–5861, 2020

  7. [7]

    Delving deep into rectifiers: Surpassing human-level performance on imagenet classification

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026–1034, 2015

  8. [8]

    Neural tangent kernel: Con- vergence and generalization in neural networks

    Arthur Jacot, Franck Gabriel, and Clement Hongler. Neural tangent kernel: Con- vergence and generalization in neural networks. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems , volume 31. Curran Associates, Inc.,

Show all 17 references
  1. [9]

    Neural tangent kernel: Convergence and generalization in neural networks

    Arthur Jacot, Franck Gabriel, and Cl ´ement Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems , 31: 8571–8580, 2018

  2. [10]

    The break-even point on optimization trajectories of deep neural networks

    Stanislaw Jastrzebski, Maciej Szymczak, Stanislav Fort, Devansh Arpit, Jacek Tabor, Kyunghyun Cho*, and Krzysztof Geras*. The break-even point on optimization trajectories of deep neural networks. In International Conference on Learning Representations, 2020. URL https://openr...

  3. [11]

    Learning curves for continual learning in neural networks: Self-knowledge transfer and forgetting, 2022

    Ryo Karakida and Shotaro Akaho. Learning curves for continual learning in neural networks: Self-knowledge transfer and forgetting, 2022. URL https://arxiv.org/abs/2112. 01653

  4. [12]

    Similarity of neural network representations revisited

    Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International conference on machine learning , pages 3519–3529. PMLR, 2019

  5. [13]

    A theory of neural tangent kernel alignment and its influ- ence on training

    Haozhe Shan and Blake Bordelon. A theory of neural tangent kernel alignment and its influ- ence on training. arXiv preprint arXiv:2105.14301, 2021

  6. [14]

    Tensor programs ii: Neural tangent kernel for any architecture, 2020

    Greg Yang. Tensor programs ii: Neural tangent kernel for any architecture, 2020. URL https://arxiv.org/abs/2006.14548

  7. [15]

    Greg Yang and Edward J. Hu. Feature Learning in Infinite-Width Neural Networks, July 2022. URL http://arxiv.org/abs/2011.14522. arXiv:2011.14522 [cs]

  8. [16]

    On the cone effect in the learning dynamics, 2025

    Zhanpeng Zhou, Yongyi Yang, Jie Ren, Mahito Sugiyama, and Junchi Yan. On the cone effect in the learning dynamics, 2025. URL https://arxiv.org/abs/2503.16316. 8 REACTIVATION : E MPIRICAL NTK D YNAMICS UNDER TASK SHIFTS Appendix A. The NTK Framework and NTK Spectrum Consider a ...

  9. [2018]

    URL https://proceedings.neurips.cc/paper_files/paper/2018/ file/5a4be1fa34e62bb8a6ec6b91d2462f5a-Paper.pdf

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.