acceptodds
Under review as a conference paper at ICLR 2027

A mechanistic theory of feature learning

Abstract

We present a theory of feature learning in overparameterized neural networks built on the dynamics of the “active parameter subspace”: the low-dimensional subspace of parameter space spanned by the network Jacobian. We show that this subspace is determined by the instantaneous gradient of the network with respect to its parameters, that it evolves throughout training, and that the curvature of the loss acts as the infinitesimal generator of its evolution. Thus, feature learning becomes a trajectory on the Grassmannian of parameter subspaces, driven by the Hessian of the loss, which in turn depends on weight decay. This perspective contrasts with the neural tangent kernel (NTK) approximation, in which learning reduces to coefficient updates within a fixed basis, and with phenomenological kinematics accounts that describe feature evolution without an underlying generator. Numerical experiments on fully connected networks trained with full-batch gradient descent support the picture: weight decay sustains the subspace trajectory after training loss has plateaued, enabling the network to discover a representation that also generalizes, while without weight decay the active subspace freezes and generalization fails. This framework may explain why delayed generalization occurs, and why weight decay is strongly associated with grokking.

Primary area:
learning theory
Keywords:
deep learning theory, intrinsic dynamics, training dynamics, neural tangent kernels, symmetric spaces, Grassmanian, grokking, generalization, feature learning, rich regime

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_zqTHOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper provides an analysis of gradient flow in neural networks, followed by identifying how the learning direction changes as time increases. The authors then comment on how weight decay can numerically cause better generalisation.

Critical strengths

S1: The paper offers a perspective on the dynamics under gradient flow, and identifies that in the overparameterized setting the instantaneous update to the weights is rank constrained.

Critical weaknesses

W1: The paper does not offer a stated theorem or result with derivation.

W2: The in-subspace rotation remark is automatic from the fact that V^T V = I

W3: The paper doesn’t reach a conclusion in the main body, but offers a lengthy introduction to gradient flow. It would be better to cut the derivation from first principles, and expand on the novel part

W4: Limited experimentation

Questions for authors

Q1: What is the main takeaway from the paper? The authors state that from the experiments weight decay increases generalisation – is there an argument for this based on the analysis in the main paper?

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

This paper proposes a geometric theory of feature learning in over-parameterized networks. Training is framed as pivoting between n-dimensional active subspaces on the Grassmannian, driven by the loss Hessian. The authors derive closed-form expressions for the infinitesimal subspace perturbation, decomposing it into in-subspace rotation (A) and out-of-subspace mixing (B, the Grassmannian velocity). Weight decay provides a persistent term in the Hessian generator that sustains B after memorization (when residuals vanish), enabling continued search for a generalizing subspace. Experiments on teacher–student MLPs with MSE loss and full-batch GD confirm that weight decay sustains the diagnostic EtE_{t} (measuring subspace rotation) and enables generalization, while without it the subspace freezes.

Critical strengths
  1. Novel geometric mechanism for grokking. Framing training as Grassmannian subspace evolution yields a concrete, testable mechanism: weight decay sustains the Hessian generator's driving term after residuals vanish. This is a refreshing alternative to phase-transition and minimal-norm explanations, and the diagnostic EtE_t provides a direct empirical handle on the proposed dynamics.

  2. Clean theoretical derivation with direct empirical support. The closed-form perturbation expressions (eqs. 13a–b), the identification of B as Grassmannian velocity, the Lie-theoretic interpretation via symmetric spaces, and the proof that Et≈∣Bt∣F/nE_t ≈ \\|B_t\\|_F/\sqrt{n} to first order form a coherent package. The teacher–student experiments directly test the theory's core predictions and are well-controlled.

Critical weaknesses
  1. Narrow empirical scope undermines the "mechanistic theory" claim. Validation is restricted to MSE loss, full-batch GD, and a single 5-100-100-5 MLP architecture. The paper acknowledges this, but the gap to practical settings (cross-entropy, mini-batching, adaptive optimizers, Transformers) is substantial. Moreover, the experiments appear to use a single random seed with no error bars or multiple runs, raising concerns about the robustness of the reported trends (especially the claim that α = 2 and α = 3 follow the "same route").

  2. The generalization argument is heuristic and not quantitatively validated. The claim that 𝔼[H] ≈ ΣpopΣ_{pop} after memorization aligns the active plane with population covariance is developed only in Appendix A without rigorous justification. More importantly, the link between EtE_{t} (subspace rotation) and test performance is correlational, not quantitative—the paper does not show how much rotation is needed for generalization, nor predict the number of steps to generalization from theory.

Questions for authors
  1. Direct test of population alignment. The paper argues that after memorization 𝔼[H] ≈ ΣpopΣ_{pop} orients the active plane toward generalizing directions. Can you provide a direct quantitative test—e.g., measuring the overlap between the active subspace Im(Vt)Im(V_t) and the top principal components of the empirical population covariance—to substantiate this claim? Without such a test, the mechanism remains a plausible narrative rather than a validated prediction.

  2. Higher-order corrections and sensitivity of EtE_t. The diagnostic EtE_t is shown to equal ∣Bt∣F/n\\|B_t\\|_F/\sqrt{n} only to first order, and the derivation assumes small learning-rate steps. How large are the 𝒪(η²) corrections in your experiments (η = 0.1)? Additionally, how sensitive is EtE_t to the checkpoint interval Δ = 100? A sensitivity analysis would clarify whether the reported plateaus are robust features or artifacts of the diagnostic's discretization.

  3. Partial generalization without weight decay. Figure 6 shows that without weight decay the model still reaches ~20% test accuracy. How does the theory explain this partial generalization? Is it due to residual-driven Hessian contributions during the memorization phase, as you suggest, and what predicts whether this partial alignment plateaus versus continues to improve? This bears directly on the claim that weight decay is the sustaining (not originating) force.

  4. Predictive power for α dependence. You observe that α = 3 takes longer but follows a similar cumulative EtE_t trajectory to α = 2. Is this a generic prediction of the theory, or an empirical observation from two values? Can the theory predict the number of steps to generalization as a function of α (e.g., from the initial distance to the teacher or the initial Grassmannian velocity)? A third value of α (e.g., α = 1.5 or 4) would significantly strengthen this claim.

Reviewer_Zw2pOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This manuscript presents a locally linearized analysis of batch-gradient-descent-based neural network training, from the perspective of explainable learning theory of neural network. This analysis is relatively straightforward in its approach and easy to comprehend its applicability and limitations. It also allows connecting related topics such as generalization, learning dynamics, and grokking. Empirical validations are limited but pointing to the core validation of the concept predicted by the theory.

Critical strengths

This is a straightforward local analytic approach to neural network training dynamics. The perspective of using evolving active subspace (and its complement space) to a small temporal vicinity of a dynamically evolving neural network is a refreshing thought. The theory developed did not tap into the Grassmannians enough, but the PoC-level demonstration of the theory in simulated learning tasks give some credibility for real-world applicability.

Critical weaknesses

There were several readability issues in this manuscript, such as typo, ineffective caption, and less-explained notation sprinkled throughout the manuscript. For example, 'without "lack" of generality', 'we interpret "asthe" trajectory', caption of Figure 3 "comparison" (of what?), the use of square brackets in the numerator of CtC_t term in Eq. (14), and normalization expression of E~t\tilde{E}_t on a line near 428. The overall reading/flow of the research could be a lot smoother without the hiccups.

Questions for authors

There seem to be untapped connections to Grassmannian arising from analyzing function and loss Hessians in neural network learning. I was repeatedly teased by multiple reference of Grassmannians, but did not get enough answered from the main text. What are your takes on this theoretical direction?

A few clarifications

  • what did you mean by the square bracketed K_{t+1} and K_t in Eq. (14)?
  • what did you mean by the angled brackets in the normalization expression of E~t\tilde{E}_t near line 428?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.