acceptodds
Under review as a conference paper at ICLR 2027

Beyond Kernel Dynamics: Label Influence at Feature Emergence.

Abstract

Linear influence estimates the effect of training-label perturbations using the derivative of a learned predictor. We show that feature emergence can sharply restrict its range of accuracy, even when this derivative is computed exactly. We study factorized models and smooth, bias-free two-layer neural networks of fixed finite width trained from small initialization by squared-loss gradient flow and fixed-step gradient descent. Near first emergence, the label Jacobian separates into a rank-one timing term proportional to and a bounded remainder. Under nondegeneracy and, for gradient descent, stability and moment conditions, relative root-mean-square linearization error under isotropic Gaussian label noise vanishes if and only if . At the critical scale, the response converges to a nonlinear random shift along the nominal trajectory. Propagated Gauss–Newton sensitivity misses the growing timing term; retaining curvature along the training direction restores it under gradient flow. We construct a derivative-preserving correction from the nominal trajectory and exact label response, prove critical-scale accuracy for exact trajectory evaluation, and give a finite-perturbation error bound for gradient flow. In a prospective test on width-64 networks, predictions from transfer to across three training configurations. A rate-based correction reduces retraining-forecast error by – at the primary emergence checkpoint and two prespecified noise levels, compared with exact linear influence. Multi-seed comparisons distinguish curvature-approximation error from finite-perturbation error and identify phase-dependent and small-noise failures.

Primary area:
learning theory
Keywords:
Influence functions, feature learning, neural network training dynamics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_rrtjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies the sensitivity of bias-free two-layer linear network predictions to training-lable perturbation when a network is initialized close to zero. The main convergence theorem (Theorem 1) showing the sensitivity is established for both GF and GD. The authors further claims that the sensitivity phenomenon can be missed by Propagated GN approximation (Section 2).

Critical strengths

Their main convergence theorem (Theorem 1) may potentially contain interesting mathematical results. However, I have concern about their interpretation, which may affect its significance.

Critical weaknesses
  1. First of all, I find the submission hard to follow due to inconsistent and unclear technical notations in addition to the lack of motivation before introducing such notations and their main theorem. For example, Section 2 introduces A_\varepsilon and t_\varepsilon together before introducing \varepsilon. There is inconsistent use of A_t and A_\varepsilon in Section 2. You have to wait until Section 3 for the formal introduction of \varepsilon, where the motivation or the role of \varepsilon is not necessarily clearly stated. It seems that all of the notations in Section 2 can be introduced without \varepsilon, and it will help fix the inconsistency issue. The paper keeps introducing technical notations in (eq-6) without explaining what they are and go straight forward to state Theorem 1. (eq-7) seems to be their main mathematical finding but its interpretation is not given. If my understanding is correct, the authors seem to claim that as \varepsilon \to 0, the emergence time T_L gets larger while the perturbation size z/L gets smaller. Thus the convergence to the profile which depends on \tau and z shows the growing sensitivity to a small label perturbation as initialization gets smaller. However, this interpretation is not immediate from the current main text.
  2. In addition to the interpretability issue, I have a substantial concern with the mathematical significance of Theorem 1. While the authors seem to claim that (eq-7) shows the sensitivity to a small perturbation, it appears to me that it is merely the effect of change-of-variable. This makes me wonder whether the sensitivity interpretation is the right way to see this equation, and wonder how general their claim could be.
  3. The authors claim in Section 4 that the phenomenon observed in Theorem 1 can be missed with the propagated GN approximation. I wonder if this is actually surprising. Comparison with the GN approximation would make sense if the R_t term is negligible, but is it really the case here? If R_t is not negligible, then it is not surprising that the GN approximation exhibit a different behavior.
Questions for authors
  1. Can you provide a very clear, informative interpretation of Theorem 1 along with more detailed explanation of technical notations (I appreciate if you could correct inconsistency among notations altogether)?
  2. Can you clarify elaborate on your sensitivity interpretation? Is it different from the effect of change-of-variable?
  3. Can you elaborate on why the GN approximation is the right object of comparison?
6 comments

Thank you for identifying the clarity problems. We will clarify the abstract, introduction, and explanation of Theorem 1.

Q1. What does Theorem 1 mean, including Eq. (7)?

The theorem concerns smooth, bias-free nonlinear two-layer networks, with both layers trained and any fixed finite width. Label changes shift when a feature is learned, so two runs evaluated at the same time can be at different learning stages.

Let ϵ\epsilon be the initialization scale and L=log⁡(1/ϵ)L=\log(1/\epsilon). Under gradient flow, the leading emergence time is TL(y)=L/λ(y)T_L(y)=L/\lambda(y), where λ(y)>0\lambda(y)>0 is the label-dependent growth rate. The transition occurs over an O(1)O(1) time interval. For δy=z/L\delta y=z/L, Eq. (7) states:

Fϵ(TL(y0)+τ;y0+z/L)=Φ(τ+q⊤z)+o(1).F_\epsilon(T_L(y_0)+\tau; y_0+z/L)=\Phi(\tau+q^\top z)+o(1).

Here Fϵ(t;y)F_\epsilon(t;y) denotes predictions at time tt for labels yy, and y0y_0 denotes nominal labels:

  • Φ\Phi is the limiting prediction curve through nominal emergence.
  • τ\tau measures time from nominal emergence.
  • z=Lδyz=L\delta y is the rescaled label change.
  • qq gives the leading advance Lq⊤δy=q⊤zLq^\top\delta y=q^\top z; a negative value means a delay.

The remainder vanishes as ϵ→0\epsilon\to0 for fixed τ,z\tau,z. The evaluation time is fixed using nominal labels, not adjusted for each perturbed run.

Linear influence replaces the shifted curve by its tangent. Under the stated assumptions and isotropic Gaussian label noise of standard deviation σ\sigma, its relative RMS error vanishes as ϵ,σ→0\epsilon,\sigma\to0 if and only if σL→0\sigma L\to0. At σ≍1/L\sigma\asymp1/L, the label change vanishes but the timing shift remains finite.

We will introduce ϵ\epsilon at the start of Section 2, distinguish the time-indexed derivative AtA_t from its checkpoint evaluation AϵA_\epsilon, and define every symbol in Eq. (6) before use.

Q2. Is this merely a change of variable?

No. The result is not merely a change of variables, for two reasons.

First, the comparison is fixed and observable. Both runs share their initialization and training time. Changing coordinates does not remove their prediction difference. Comparing runs at their own emergence times changes the question; under the stated retiming conditions, the order-LL term cancels and the response remains bounded (Eq. 15, Appendix E).

At L=40L=40, phase 00, and critical noise level c=1c=1, Table 1 reports median fresh-output relative RMS errors of 0.3740.374, 0.4300.430, and 0.1280.128 for exact linear influence in B1–B3, versus 0.0880.088, 0.0880.088, and 0.0250.025 for the rate-clock correction, over five paired initialization seeds. B1 matches the theorem's bias-free architecture; B2 and B3 train biases and test beyond that assumption.

Second, we prove the nonlinear emergence profile and control its label derivatives. The chain-rule calculation is elementary once a sufficiently regular profile exists. The required proof establishes:

  • Uniform control of the profile's dependence on nearby labels.
  • Label-derivative estimates over the growing training interval, yielding an order-LL response and bounded remainder.
  • Bounds establishing Gaussian mean-square limits and both directions of the σL\sigma L boundary.

Appendices A.1 (Lemma 4) and A.2 establish derivative control and the response expansion; Appendix B proves the noise boundary. A nominal trajectory limit alone does not justify differentiation with respect to labels.

First emergence is covered for gradient flow and fixed-step GD; GD's Gaussian statement requires stability and moment conditions. Later-transition results require the diagonal model's separated activations and no deactivations; they do not cover general MLPs.

Q3. Why GN? Is residual curvature negligible?

No. It is not negligible for the leading response. During the long small-output phase, residuals approach −y0-y_0; in this model, residual curvature generates the unstable growth responsible for timing amplification. We ask: does following evolving features suffice to recover that response?

Propagated GN uses the same nominal trajectory and label forcing as exact sensitivity, including every feature update. Under the theorem's assumptions, exact sensitivity is Θ(L)\Theta(L), whereas positive-semidefinite curvature propagation in this setup remains O(1)O(1). GD additionally requires stable propagation. Table 1 gives GN errors of 0.9590.959–0.9610.961 across B1–B3; predicting no change gives 11.

This curvature choice is relevant to attribution: SOURCE uses approximate unrolling with an EK-FAC approximation to Gauss–Newton curvature (Bae et al., NeurIPS 2024, Section 3.3).

Corollary 2 and Appendix D.1 identify a sufficient repair. Under gradient flow, a rank-at-most-two correction retaining residual-curvature action along parameter velocity recovers the leading order-LL response with bounded error remaining. Its action requires a residual-Hessian–vector product.

AuthorsOpenReview ↗

Revision uploaded: interpretation of Theorem 1 and the GN comparison

Thank you for identifying the notation and presentation problems and for questioning the interpretation and significance of Theorem 1. We have uploaded a revised manuscript addressing each of these concerns.

1. The theorem and its notation are now explained before use

The abstract and introduction now explain the prediction question, linear influence, and feature emergence in words. The introduction defines the initialization scale ϵ\epsilon and L=log⁡(1/ϵ)L=\log(1/\epsilon). Section 2 distinguishes the time-dependent derivative Aϵ(t;y)A_\epsilon(t;y) from its checkpoint value AϵA_\epsilon. Section 3 defines the growth rate, emergence time, phase, and timing direction before Theorem 1, then adds “Reading the result” and “Why the nonlinear-network proof is needed.” Figure 1 illustrates the mechanism.

The theorem concerns smooth, bias-free nonlinear two-layer networks, with both layers trained and any fixed finite width, under gradient flow or fixed-step GD. The solvable scalar model is an illustration.

2. The sensitivity is a fixed-time prediction difference, not a coordinate artifact

No, the result is not merely a change of variables. Both runs start from the same parameters and are evaluated after the same training time. Changing the horizontal axis of a plot cannot remove the difference between the predictions recorded at that checkpoint. Stopping each run at its own learning stage changes the experiment.

Here is the distinction in the theorem. Let Φ(τ)\Phi(\tau) denote the limiting prediction curve through emergence, with τ\tau measuring time relative to the nominal transition. A label perturbation produces a timing advance dd, which remains finite even as that perturbation vanishes at the critical scale of order 1/L1/L. As initialization tends to zero, the actual change and linear forecast have the limits

\text{linear forecast}\ \longrightarrow\ \Phi'(\tau)d.$$ Linear influence follows the tangent; retraining follows the curve. Under the theorem's conditions, this distinction gives a strictly positive limiting relative RMS error for training predictions at the critical Gaussian noise scale. It is a statement about observable predictions in the original training procedure. The proof uses normalized time to control the dynamics, but the theorem returns to the original, common observation time. The difficult step is proving that this return is valid for both the nonlinear trajectory and its label derivatives. Ordinary smooth dependence over a fixed time interval does not supply uniform bounds over the waiting period, which grows as initialization shrinks. Lemma 4 establishes the required control; Appendix A.2 derives the growing timing term and bounded remainder; Appendix B establishes the necessary-and-sufficient noise boundary. Thus the result does not follow by applying the chain rule to an assumed limiting curve. The alternative observation rule is explicit: under the matched-progress conditions, allowing each run a different stopping time cancels the leading timing term (Eq. (11), Appendix E). **3. Why the GN comparison is informative** Residual curvature is **not negligible** here. During the long initial small-output phase, the network has not fitted the labels, and its residual approaches the negative label vector. Section 4 now states this directly. The point is not that omitting a non-negligible term can change an answer. We identify **which part of the response is lost, how its size scales, and a sufficient way to recover it**. Propagated GN follows the same nominal training trajectory and retains every feature update and the same label forcing. Under the theorem's conditions, the positive-semidefinite propagation in Section 4 remains $O(1)$ at emergence, while the exact response grows as $\Theta(L)$. At critical noise of order $1/L$, the GN prediction therefore vanishes while the actual prediction change remains nontrivial. Its relative RMS error tends to one, the error of predicting no change. The discrete result requires the stable-propagation conditions stated in Appendix D. This separation has a constructive counterpart: under gradient flow, Corollary 2 retains residual-curvature action along the parameter velocity using a correction of rank at most two and recovers the leading response with bounded error. It identifies a sufficient curvature component to restore, rather than requiring the full residual Hessian. In bias-free B1, the median operator norm of the exact response divided by $L$ changes from 1.28 to 1.23 as $L$ increases from 5 to 40, while the unscaled GN norm stays between 1.59 and 1.65 (Section 6). For reference, original Eq. (7) is now Eq. (5), original Eq. (15) is Eq. (11), and original Table 1 is Table 2. Theorem 1 and Corollary 2 retain their numbering. We would appreciate your assessment of these clarifications.

Thanks for your revised manuscript and I appreciate the authors effort. However, the revision contains extensive changes, including substantial rewriting and restructuring as well as noticeable shift in the positioning of the paper. While I appreciate the authors effort to clarify our concerns, I am not sure reviewers are expected to conduct a new review of the revised manuscript during this short period.

As for my concerns,

  1. I don't think the authors made explicit my interpretation of equation (5) is correct or not directly.
  2. I was not asking whether the proof requires technically difficult argument or not. I was asking whether the mechanism can be largely reduced to cancellation of scaling: T_L scales by L while z/L scales by 1/L and they cancels out. It seems to me the equation TL(y0)−TL(y0+δy)=Lq⊤δy+O(L∥δy∥2)T_L(y_0) - T_L(y_0 + \delta y) = Lq^\top \delta y + O(L\|\delta y\|^2) preceding Theorem 1 reinforces my point. Since the network remain close to zero until it reaches near tϵt_\epsilon, I'm not sure if this is something to be surprised at. Again I want to make it clear that I'm not saying proving Theorem 1 is easy. I recognize the technical difficulty of proving it rigorously. My concern is solely about the significance. "Sensitivity is a fixed-time prediction difference" seems to reinforce my point. It seems to me the choice of fixed time tϵ=TL(y0)+τt_\epsilon = T_L(y_0) + \tau is the key of equation (5).
  3. I think your correction made my point for the propagated GN approximation. Since RtR_t is not negligible, it is not surprising the approximation exhibit a different behaviour.
Reviewer_BGtAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper shows that label perturbations shift feature emergence in two-layer networks trained from small initialization, which limits linear influence even with exact derivatives. It characterizes these limits and proposes trajectory-based corrections.

Critical strengths

The paper distinguishes derivative-approximation error from the limitations of linear prediction under label perturbations. Using two-layer networks trained from small initialization, it characterizes the feature-emergence shift as a concrete failure mechanism and establishes a sharp validity boundary for exact linear influence. The analysis also motivates derivative-preserving trajectory corrections, a valuable contribution supported by theoretical guarantees and controlled experiments.

Critical weaknesses

Presentation

I found the paper very difficult to follow, beginning with the abstract and introduction. For example, the abstract introduces L=log⁡(1/ϵ)L = \log(1/\epsilon) and the condition σL→0\sigma L \to 0 without explicitly explaining what ϵ\epsilon and σ\sigma represent. It also refers to a “rank-one timing term,” the “critical scale,” and predictions transferring from L=20L = 20 to L=40L = 40 without enough context to understand them on a first reading. These quantities and concepts are explained later, but that does not make the initial presentation accessible. I encountered this pattern repeatedly in the main text, where understanding a claim often required reading ahead and then revisiting earlier passages.

The transitions between results also make the main argument difficult to track. Section 4 moves from correcting curvature-based derivative approximations to matched-progress retiming, before Section 5 returns to forecasting finite label changes at a fixed checkpoint. These address different questions: improving a derivative estimate, changing what is compared, and improving finite-perturbation predictions. Although the paper acknowledges these distinctions, it could explain their roles in the overall argument more clearly before presenting the individual results. The separate diagonal-network extension in Section 6 also interrupts the progression from the forecasting method in Section 5 to its evaluation in Section 7.

A more self-contained abstract, clearer explanations of key quantities at first use, and a roadmap distinguishing the central argument from its consequences and extensions would substantially improve readability.

Technical Novelty

The paper acknowledges prior work on stagewise influence, but the technical advance beyond existing solvable models could be made more explicit. Clarifying which conclusions follow from known emergence dynamics and which require the new uniform label-response analysis would help readers better assess the contribution.

Questions for authors

Could you clarify which conclusions of Theorem 1 can already be obtained in solvable deep-linear models and which additional estimates are needed for the nonlinear finite-width setting? A focused comparison with the analytical treatment in Lee et al. (2026) would be helpful.

6 comments

Thank you for identifying where the presentation makes the argument difficult to follow. We agree that the central question, notation, and relation to prior work should be clear without reading ahead. We clarify them below and describe the revisions we plan; the submitted PDF has not yet been revised.

1. The main argument and a clearer roadmap.

The question is: how large a label change can the exact derivative reliably predict at a fixed training checkpoint? A derivative can be computed accurately while its linear prediction of a finite change is inaccurate.

Here ϵ\epsilon is the initialization scale, σ\sigma is the standard deviation of additive Gaussian label noise, and L=log⁡(1/ϵ)L=\log(1/\epsilon). Near first emergence, a label change advances or delays learning. The dominant part of the derivative points along the velocity of the nominal prediction curve. This is the meaning of the "rank-one timing term." Noise of size 1/L1/L can produce a time shift comparable to the duration of the transition, even as the label change tends to zero. This is the "critical scale."

The main sequence is fixed-checkpoint sensitivity and its validity boundary, then a forecast for finite label changes, then experiments. Three related results have distinct purposes:

  • Curvature repair improves the derivative estimate.
  • Matched-progress evaluation compares runs at corresponding learning stages, changing the counterfactual.
  • Trajectory forecasting improves finite-change prediction at the original, common checkpoint.

Our planned revision makes this sequence explicit, presents the first two as separate consequences, and moves the diagonal-network extension so it does not interrupt the forecast-to-experiment sequence. We will make the abstract self-contained, end the introduction with a roadmap, and define every quantity at first use. The L=20L=20 to L=40L=40 experiment will be introduced as forecasting across initialization scales, rather than as an unexplained numerical comparison.

2. What is new beyond Lee et al.?

Lee et al. (ICLR 2026, Appendix C.6) differentiate finite-time deep-linear trajectories under data upweighting to study how influence changes during learning. We establish when the exact derivative reliably predicts finite label changes in nonlinear finite-width networks. The distinction concerns both the question and the mathematical setting.

Explicit scalar trajectories already reveal logarithmic emergence times, growing sensitivity, and nonlinear timing shifts. Our contribution is to establish a sharp finite-perturbation law for smooth, bias-free two-layer networks with both layers trained and any fixed finite width, without relying on an explicit solution. Under the stated assumptions, as initialization and noise vanish, relative RMS linearization error vanishes if and only if σlog⁡(1/ϵ)→0\sigma\log(1/\epsilon)\to0. At the critical scale, we characterize the nonlinear prediction response and prove that a derivative-preserving trajectory correction is asymptotically accurate.

The essential additional step is uniform control of the nonlinear emergence trajectory and its label derivatives over a training horizon growing as log⁡(1/ϵ)\log(1/\epsilon). These estimates justify differentiating the limiting trajectory and control the remainder. Neither nominal emergence dynamics alone nor the chain rule supplies those guarantees.

3. The additional estimates behind Theorem 1.

The proof requires three controls:

  • Uniform dependence on nearby labels. Lemma 4 (Appendix A.1) controls the emergence trajectory and its first two derivatives in labels and phase over the growing training interval. Knowing the trajectory for one set of labels does not justify differentiating its limiting form.
  • A bounded remainder. Appendix A.2 shows that the exact label derivative consists of a timing component proportional to LL and a bounded remainder. The chain rule gives the formal decomposition; the estimates justify it asymptotically and bound the remainder throughout the transition.
  • Control of random perturbations. A local trajectory limit does not by itself establish accuracy averaged over unbounded Gaussian noise. Appendix B supplies the additional bounds needed for both directions of the validity boundary for training outputs. At the critical scale, linear influence retains a nonzero relative error. Fixed-step GD requires a separate discrete entrance construction and the additional stability and moment conditions in Appendix C for the Gaussian statement.

We will make this distinction between the solvable illustration and the nonlinear theorem explicit alongside Theorem 1 and in the comparison with Lee et al.

AuthorsOpenReview ↗

Revision uploaded: clearer presentation and technical contribution

Thank you for identifying the presentation problems and for asking us to distinguish the technical advance from the solvable-model analysis in Lee et al. We have uploaded a revised manuscript addressing both concerns.

1. What we changed to make the argument easier to follow

  • The abstract now explains the prediction question, linear influence, and feature emergence in words before stating the result. It no longer relies on an unexplained LL or initialization-scale comparison.
  • The introduction motivates the question through uncertain or corrected training labels, defines the key quantities, and ends with a roadmap.
  • Section 3 introduces the notation before Theorem 1, separates its conclusions, and follows it with “Reading the result” and “Why the nonlinear-network proof is needed.” Figure 1 illustrates why a finite timing shift differs from a tangent prediction at the same checkpoint.
  • We distinguish three operations: improving the derivative estimate, evaluating runs at matched learning progress, and forecasting a finite label change at a common checkpoint. Forecasting in Section 5 now leads directly to experiments in Section 6; the later-transition extension follows in Section 7.

2. What the nonlinear theorem adds beyond solvable trajectories

Lee et al. (ICLR 2026, Appendix C.6) differentiate finite-time deep-linear trajectories to study influence during stagewise learning. Logarithmic waiting times, amplified derivatives, and nonlinear timing shifts can already be seen in solvable models. Our scalar example makes this mechanism explicit.

Our main contribution is a sharp limit on how far the exact derivative remains predictive, proved for nonlinear networks without a closed-form training solution.

The distinction is concrete. A large derivative measures the slope at the original labels. It does not tell us how far that slope remains an accurate forecast. Near emergence, a small label change moves the network along a curved part of its learning trajectory. The derivative gives the tangent displacement; retraining follows the curve. Both predictions are evaluated at the same training time.

For smooth, bias-free two-layer networks with both layers trained and any fixed finite width, Theorem 1 establishes the boundary between these behaviors. Writing ϵ\epsilon for initialization scale and σ\sigma for isotropic Gaussian label-noise standard deviation, relative RMS linearization error for training predictions vanishes precisely when σlog⁡(1/ϵ)→0\sigma\log(1/\epsilon)\to0, under the stated assumptions. At noise proportional to 1/log⁡(1/ϵ)1/\log(1/\epsilon), we identify the nonlinear limiting response. For gradient flow, Proposition 3 also gives a finite-perturbation error bound establishing the correction's accuracy at that scale.

The additional mathematical work is what makes these conclusions valid beyond an explicit scalar solution:

  • Control through the nonlinear transition. The initial exponential growth explains the waiting time but ceases to describe training once the network leaves the small-output state. Lemma 4 controls the nonlinear continuation and its first two label derivatives uniformly over the growing waiting period.
  • Control of the derivative remainder. Appendix A.2 turns those estimates into a growing timing term plus a bounded content response. Convergence of nominal trajectories alone would not justify differentiating their limit.
  • Control of random perturbations. Appendix B supplies the bounds needed to pass from local trajectory estimates to Gaussian mean-square accuracy and to establish both directions of the boundary. Fixed-step GD uses a separate discrete construction, with the additional stability and moment conditions stated in Appendix C.2.

The introduction, the explanation following Theorem 1, and Section 8 now make this distinction explicit.

3. Evidence accompanying the revised explanation

Revised Table 1 and Appendix I.9 also incorporate the five independently sampled datasets described in our responses. At the primary emergence checkpoint and two primary noise levels, the rate correction improves on exact linear influence, propagated GN, and held-out shrinkage in every dataset and branch, using median paired comparisons. This supports the constructive consequence of the theory in the tested width-64 networks.

Theorem 1 and Lemma 4 retain their numbering; the fixed-checkpoint law is now Eq. (5), previously Eq. (7). Our top-level announcement supplies the remaining reference changes. We would appreciate your assessment of whether the revised exposition and explicit comparison resolve the concerns you raised.

I appreciate the authors’ efforts in their response and revision. Most of my concerns have been resolved. However, I still have some concerns about the presentation. Although the revised manuscript is more complete than the previous version, its flow and clarity could be improved further. I encourage the authors to improve the clarity of the paper. That said, I am happy to lean toward a positive recommendation.

Reviewer_jqXROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies how far label influence—the derivative of a model's predictions with respect to training labels at a fixed training checkpoint—remains accurate for finite label changes near feature emergence. For smooth, bias-free two-layer networks trained from small initialization, the label Jacobian near first emergence decomposes into a rank-one term of order L=log⁡(1/ϵ)L=\log(1/\epsilon), reflecting label-dependent transition timing, plus a bounded remainder (Theorem 1). Under the stated assumptions, the relative RMS linearization error under Gaussian label noise vanishes if and only if σL→0\sigma L\to0. The authors analyze gradient flow and fixed-step gradient descent, show that propagated Gauss–Newton sensitivity misses the growing timing term, and prove a low-rank curvature repair under gradient flow. They also propose a derivative-preserving forecast that applies the timing shift nonlinearly along the nominal trajectory. Experiments on width-64 tanh networks at L=5,10,20,40L=5,10,20,40 support finite-scale consequences of the theory and show large forecast improvements near emergence at the primary noise levels.

Critical strengths
  1. The question is well posed: the paper distinguishes the accuracy of the exact derivative as a finite-perturbation forecast from the accuracy of estimating that derivative. The identified linearization failure cannot be removed merely by estimating the same derivative more accurately. The quantitative validity boundary sharpens the observation that influence estimates can be fragile.

  2. The empirical protocol is careful and transparently reported: predictions for the target initialization scale L=40L=40 were fixed before those runs, a held-out shrinkage baseline is included, forecasts exceeding the continuation cap are marked unavailable, and negative results are retained.

  3. The proposed forecast preserves the exact local derivative by construction (Eq. 16), retaining first-order agreement with exact influence. The reported paired error reductions of 70–83% for the rate correction at the primary checkpoint and the two prespecified noise levels are substantial.

Critical weaknesses
  1. Empirical scope is limited, as the authors acknowledge: one synthetic data realization, five initialization replicates, a single architecture family (width-64, two-layer tanh), and squared-loss regression. It is therefore not established how the mechanism behaves under independently sampled data or in larger-scale training.

  2. Two of the three configurations in the prospective experiment, B2 and B3, train biases and fall outside the bias-free assumption of Theorem 1. Appendix I.4 supplies a centered analytic clock for these branches, but does not establish a corresponding theorem. Their results provide empirical robustness evidence; the theoretical conditions for extending the decomposition to trained biases remain unresolved.

  3. Practical applicability is regime-dependent. At the subcritical setting c=0.3/L=0.0075c=0.3/L=0.0075 for L=40L=40, the rate correction beats exact linear influence in only 13 of 105 repeated comparisons (Table 2). For a single forecast, the full label Jacobian is unnecessary, but the method still requires a directional label response through training, access to the nominal trajectory, and possibly additional nominal continuation. The authors appropriately describe it as a controlled mechanism test rather than a scalable replacement for attribution methods.

Questions for authors
  1. Given the correction's gains at the primary critical noise levels and its frequent losses at subcritical noise, could the authors suggest a criterion based on the nominal run and proposed perturbation for deciding when to apply it? Proposition 3 uses constants involving nearby-label trajectories; can a practical diagnostic account for trajectory curvature and interpolation error without perturbed retraining?

  2. The geometry measurements and centered analytic clock provide empirical support for B2 and B3. Could the authors give a theoretical justification for that clock and identify conditions under which the rank-one timing decomposition extends to trained biases, beyond the robustness evidence currently reported?

  3. Appendix B already covers correlated Gaussian noise with q⊤Σq>0q^\top\Sigma q>0. What linearization boundary should be expected when structured perturbations are orthogonal or nearly orthogonal to the timing direction, including covariance supported in ker⁡q⊤\ker q^\top? Could the authors characterize how higher-order clock changes and the bounded content response determine the relevant scale in this excluded case?

3 comments

Thank you for the constructive review. We have addressed the single-dataset concern with new experiments.

Independent-data replication. We repeated the fixed protocol on five independently sampled datasets, each with five initialization seeds, across B1–B3: 75 nominal runs and 6,000 perturbed trajectories. At phase 0 and c=0.3,1c=0.3,1, the rate correction beat exact linear influence, propagated GN and held-out shrinkage on every dataset in every branch, with complete primary coverage. Dataset-wise median paired error reductions versus exact linear influence ranged from 67.5% to 94.8%. This establishes replication across data draws; the architecture, loss and data-generating process remain fixed.

Q1. When should the correction be used?

Retain linear influence when the nonlinear correction is too small to resolve reliably. At c=0.0075c=0.0075, the correction beats linear in 69/525 comparisons across the prescribed checkpoints, versus 13/105 originally. Both forecasts agree to first order, so higher-order terms and numerical accuracy determine their small-noise differences.

For a proposed label change uu, compute the linear forecast a=Aua=Au, rate forecast PdP_d and correction size D=∥Pd−a∥D=\|P_d-a\|. Using only the nominal run, check shifted-time coverage, compare interpolation choices on the same trajectory, and measure content drift V=∥(K(τ+d)−K(τ))u∥V=\|(K(\tau+d)-K(\tau))u\|. Here dd is the rate-clock shift and KK is the finite-initialization derivative after subtracting the timing component. Computing VV requires directional sensitivity along the nominal trajectory, not perturbed retraining. Each interpolant must use the same slopes as its correction.

The decision has a precise sufficient condition: if ∥Δ−Pd∥≤B\|\Delta-P_d\|\le B for the true change Δ\Delta, then D>2BD>2B guarantees that the correction beats linear influence. Proposition 3's gradient-flow argument gives B=12Myy∥u∥2+V+IB=\tfrac12M_{yy}\|u\|^2+V+I, where MyyM_{yy} bounds nearby-label second derivatives and II bounds numerical error in the complete forecast. Thus correction size, content drift and interpolation sensitivity give a nominal-run diagnostic. A certified selector additionally needs valid bounds on MyyM_{yy} and II; we do not claim that the experiments validate this selector.

Q2. Why does centering give the clock for trained biases?

The centered clock is the instability rate after the output bias fits the label mean. At the corresponding zero-feature equilibrium, let P=I−11⊤/nP=I-\mathbf1\mathbf1^\top/n and X~=[X,1]\widetilde X=[X,\mathbf1]. The feature growth rate is

λ~=ϕ′(0)∥X~⊤Py/n∥.\widetilde\lambda=\phi'(0)\|\widetilde X^\top Py/n\|.

The output-bias mode has eigenvalue −1-1, while GD's unstable multiplier is 1+hλ~1+h\widetilde\lambda. This yields the logarithmic rate in Appendix I.4. Along the zero-feature trajectory, the output bias approaches the mean exponentially, with 0<h<20<h<2 for GD. This transient changes the incoming unstable coefficient, but not the leading waiting time L/λ~L/\widetilde\lambda for flow or L/log⁡(1+hλ~)L/\log(1+h\widetilde\lambda) for GD, when that coefficient is nonzero.

The rank-one extension requires a uniformly positive centered rate, a nonzero incoming unstable component, uniform label-derivative control through this transient, and the stated GD stability/moment conditions. The missing proof step is adapting Lemma 4 to this transient. B2/B3 therefore test a dynamically justified clock, while remaining outside the submitted bias-free theorem.

Q3. What happens for noise orthogonal to the timing direction?

Removing the first-order timing shift does not remove higher-order timing effects. For training outputs under gradient flow, write g0=X⊤y0/ng_0=X^\top y_0/n, s=ϕ′(0)s=\phi'(0) and let centered Gaussian ZZ satisfy q⊤Z=0q^\top Z=0. Expanding the analytic clock gives

dL(σZ)=Lσ2b(Z)+O(Lσ4∥X⊤Z/n∥4),b(Z)=∥X⊤Z/n∥22s∥g0∥3.d_L(\sigma Z)=L\sigma^2b(Z)+O(L\sigma^4\|X^\top Z/n\|^4),\qquad b(Z)=\frac{\|X^\top Z/n\|^2}{2s\|g_0\|^3}.

With limiting content derivative KK and Lσ2→0L\sigma^2\to0, the leading response combines linear content σKZ\sigma KZ and quadratic timing Lσ2Φ′(τ)b(Z)L\sigma^2\Phi'(\tau)b(Z). If both coefficients have nonzero mean-square size, relative linearization error is already nonnegligible at σ≍L−1\sigma\asymp L^{-1}, although an order-one clock shift requires σ≍L−1/2\sigma\asymp L^{-1/2}. Gaussian symmetry makes these two terms orthogonal in mean square.

If X⊤Z=0X^\top Z=0 almost surely, the clock is constant: with nonzero content derivative, the usual O(σ)O(\sigma) relative-error bound applies without an LL restriction. Vanishing content requires higher-order analysis. For nearly orthogonal noise, retain both clock orders and compare them with content and trajectory curvature; q⊤Σqq^\top\Sigma q alone is insufficient.

We will incorporate the new results, diagnostic and derivations into the revision. If accepted, we also plan to evaluate a second smooth activation for the camera-ready version.

AuthorsOpenReview ↗

Revision uploaded: independent-data results and additional analyses

Thank you for highlighting the need for replication across independently sampled datasets and for your questions about when the correction is useful and how far the theory extends. We have revised the manuscript to address these points:

  • Independent-data replication: Section 6, Table 1, and Appendix I.9 now report five independently sampled datasets, with five initialization seeds per dataset in each of B1–B3: 75 nominal runs and 6,000 perturbed trajectories. At the primary emergence checkpoint and both primary noise levels, the rate correction improves on exact linear influence, propagated GN, and held-out shrinkage in every dataset and branch, using median paired comparisons. Dataset-wise median paired error reductions versus exact linear influence range from 67.5% to 94.8%, with complete primary coverage. Small-noise losses and numerical checks are also reported.
  • When to use the correction: Section 5 and Appendix F.1 present the nominal-run diagnostic and sufficient comparison bound, distinguishing what can be checked from one trajectory from what requires additional bounds.
  • Trained biases: Appendix C.3 derives the centered growth rate and identifies the additional conditions needed to extend the emergence theorem.
  • Orthogonal perturbations: Appendix B.1 derives the quadratic timing effect and explains its interaction with the bounded content response.

These additions clarify both the evidence supporting the correction and the limits of its applicability. The revision retains the distinction between the bias-free theorem and the empirical evidence from the biased configurations.

Our top-level revision announcement provides the mapping between the original and revised equation and table numbers.

We would appreciate your reassessment of the paper in light of these additional results and clarifications, and welcome any remaining questions.

Thanks for author's effort, most of the concern have been addressed. I would maintain my score.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revised manuscript: changes and reference guide

We have uploaded a revised manuscript incorporating the clarifications, additional analyses, and independent-dataset experiments discussed in our responses. The central research question and main theoretical conclusions are unchanged, and the original experimental results are retained.

The principal changes are:

  1. Clearer motivation and notation. The abstract and introduction now explain linear influence and feature emergence before introducing the technical results. We define the initialization scale and other quantities at first use, distinguish the time-dependent label derivative from its checkpoint evaluation, and provide a roadmap and a mechanism schematic.

  2. A more accessible explanation of Theorem 1 and its contribution. Section 3 separates the fixed-checkpoint prediction law, the derivative decomposition, the noise boundary, and the GD extension. The explanation makes explicit that both runs use the same initialization and observation time. It also identifies the proof requirements beyond the chain rule: uniform control of the nonlinear trajectory and its label derivatives over the growing waiting period. The introduction and Section 8 clarify the distinction from Lee et al., including their differentiation of finite-time deep-linear trajectories.

  3. Clearer organization of the comparisons. Section 4 explains why propagated GN omits the leading timing response despite following the evolving features. We distinguish derivative approximation, matched-progress evaluation, and finite-change forecasting. The forecasting method in Section 5 now leads directly to the experiments in Section 6. The extension to later transitions appears separately in Section 7.

  4. Independent-dataset evidence. Table 1 and Appendix I.9 incorporate the five-dataset study reported in our response: 75 nominal runs and 6,000 perturbed trajectories. At the primary emergence checkpoint and two primary noise levels, the rate correction improves on exact linear influence, propagated GN, and held-out shrinkage on every dataset in every branch, using dataset-level median paired comparisons. Small-noise losses and incomplete forecasts are also reported. The architecture, loss, and data-generating process remain fixed across this replication.

  5. Additional analyses addressing the questions raised. Appendices F.1, C.3, and B.1 respectively present the nominal-run diagnostic, the centered-clock calculation for trained biases, and the analysis of perturbations orthogonal to the timing direction. The text distinguishes the diagnostic from a certified selection rule and the local biased-network calculation from the bias-free emergence theorem.

Reference guide. Equation and table references in our November 11 responses refer to the original submission. The corresponding locations in the revised manuscript are:

ResultOriginal submissionRevised manuscript
Sensitivity evolutionEq. (4)Eq. (12)
Growth rate and emergence clockEq. (6)Eq. (4)
Fixed-checkpoint prediction lawEq. (7)Eq. (5)
Timing/content derivative decompositionEq. (8)Eq. (6)
Critical Gaussian responseEq. (9)Eq. (7)
Matched-progress comparisonEq. (15)Eq. (11)
Derivative-preserving forecastEq. (16)Eq. (15)
Original single-dataset forecast errorsTable 1Table 2
Original noise-sweep comparison countsTable 2Table 3

The numbering of Theorem 1, Corollary 2, Proposition 3, and Lemma 4 is unchanged. The new Table 1 summarizes the independent-dataset results.

We thank the reviewers for the questions and suggestions that helped us make the argument and its scope clearer.

AuthorsOpenReview ↗

Second revised manuscript: presentation changes and reference guide

Revised manuscript: presentation changes and reference guide

We thank the reviewers for pointing out where the manuscript required readers to look ahead. This revision focuses on rewording and reorganizing the presentation. It preserves the mathematical results, proofs, and experiments reported in our November 15 update. The changes include:

  1. Abstract and introduction. The abstract now explains how the rate-based forecast uses an estimated advance or delay in feature emergence to predict the effect of changing the labels. The introduction distinguishes the problem of estimating a label derivative from the problem of forecasting a finite label change at a fixed checkpoint. It also identifies the separate question of what response remains when runs are compared at equal learning progress.

  2. Theorem 1. Section 3 explains how the initial growth rate gives rise to the waiting time in Equation (4). The theorem now states its main conclusion before its three labeled claims. It defines the limiting prediction curve, distinguishes a shared checkpoint from a fixed offset relative to each run’s own emergence time, and names the critical noise regime at first use. Figure 1 now follows the discussion of the theorem.

  3. Derivative estimation and forecasting. Section 4 clarifies that Gauss–Newton approximates label sensitivity along the original training run. Section 5 explains that both runs are evaluated at the same checkpoint and that the time shift is used only to construct the forecast. We condensed the main-text discussion of matched progress and direct readers to Appendix E for the calculation.

  4. Experiments. Section 6 now states what each study tests before introducing the protocol. The first study uses measurements at one initialization scale to predict how much label sensitivity lies perpendicular to the direction of prediction change at another scale. The second compares forecasts with the prediction changes observed after retraining on five independent datasets. Table 1 now identifies the emergence checkpoint and the normalized label-noise levels in its caption.

Reference guide. The table maps the locations cited in our earlier responses and November 15 announcement to the current manuscript.

\textbf{Result} & \textbf{Original} & \textbf{Nov. 15} & \textbf{Current} \\\\ \hline \text{Sensitivity evolution} & \text{Eq. (4)} & \text{Eq. (12)} & \text{Eq. (11)} \\\\ \text{Growth rate and emergence time} & \text{Eq. (6)} & \text{Eq. (4)} & \text{Eq. (4)} \\\\ \text{Fixed-checkpoint prediction law} & \text{Eq. (7)} & \text{Eq. (5)} & \text{Eq. (5)} \\\\ \text{Timing and content decomposition} & \text{Eq. (8)} & \text{Eq. (6)} & \text{Eq. (6)} \\\\ \text{Critical Gaussian response} & \text{Eq. (9)} & \text{Eq. (7)} & \text{Eq. (7)} \\\\ \text{Matched-progress comparison} & \text{Eq. (15)} & \text{Eq. (11)} & \text{App. E, Eq. (43)} \\\\ \text{Derivative-preserving forecast} & \text{Eq. (16)} & \text{Eq. (15)} & \text{Eq. (14)} \\\\ \text{Five-dataset forecast comparison} & \text{—} & \text{Table 1} & \text{Table 1} \\\\ \text{Single-dataset forecast errors} & \text{Table 1} & \text{Table 2} & \text{Table 2} \\\\ \text{Noise-sweep comparison counts} & \text{Table 2} & \text{Table 3} & \text{Table 3} \end{array}$$

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.