acceptodds
Under review as a conference paper at ICLR 2027

The Memorization Cost of Generalization: A Quantitative Causal Law in Noisy-Trained Deep Networks

Abstract

Deep networks trained on noisy labels do two things at once: they learn generalization from the clean samples and they account for the noise samples by memorizing them. We show that these two processes occupy complementary channels of the representation's topology. Using a two-level persistence pipeline (Topo) that separates persistent homology into a within-class component (loops internal to a single class, the generalization channel) and a cross-class component (loops that close only across classes, the memorization channel), we establish that memory and generalization are causally separable: memory can be added, stripped, and — the central result — its cost on generalization can be billed quantitatively. Along the overlay path (freeze a clean-trained model, then unfreeze flipped samples with clean samples still present), the memorization cost obeys the law i.e., each 1% of memorized flipped samples costs validation accuracy, with the slope proportional to the noise:clean sample ratio . At the shared 11M reference capacity, the slope constant clusters near across datasets and class counts: (CIFAR-10), (SVHN), (CIFAR-100, 100 classes), and (VGG on SVHN); the per-setting spread is the image term of the displacement mechanism, not a capacity signature. The law also holds on real human label noise: on CIFAR-10N the calibrated constant predicts the measured memorization slope to a mean ratio of (3 seeds), while maximally class-confusable flips bound the coefficient's noise-structure sensitivity at . Across capacities varies systematically (0.470.34 over a width sweep) and is traced to a normalized clean-sample feature displacement, (Zhang et al., 2026, arXiv:2608.30487). Every calibration cell is 3-seed. The constant is a through-origin fit; a low-noise extrapolation to stays within on both datasets but with larger residual scatter, delimiting the law's cleanest domain to the mid-noise range . Memory, we have shown, is an invertible, stackable layer, not an uninterpretable black-box capacity: its cost to generalization obeys a quantitative causal law with a calibrated constant that clusters near across datasets and class counts at the shared reference capacity, with its capacity dependence mechanistically traced.

Primary area:
interpretability and explainable AI
Keywords:
memorization, generalization, label noise, persistent homology, topological data analysis, representation geometry, causal intervention, learning dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_xFJxOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates the relationship between memorization and generalization in networks trained with noisy labels. Using the Topo^2 framework, the authors interpret within-class and cross-class topological structures as generalization and memorization channels. They introduce overlay and stripping interventions and propose a quantitative relationship between memorization and validation-accuracy loss, with a coefficient near 0.38 in several reference settings

Critical strengths
  1. The paper addresses an interesting question: how much generalization is lost when a network fits incorrect labels.
  2. The FM0 reference and controlled exposure to flipped samples provide a useful experimental setting for investigating this trade-off.
  3. The authors report multi-seed calibration and explicitly acknowledge limitations involving model capacity, noise structure, and noise range.
Critical weaknesses
  1. The claim of causal separability is not sufficiently established. Section 3.4 states that the generalization channel is frozen, but does not specify what is actually fixed. Continuing training with noisy samples does not itself demonstrate selective intervention on an independent memorization channel.
  2. The topological channels lack precise operational definitions. It is unclear how within-class and cross-class are distinguished, and how the interpolation in Section 3.2 is performed and motivated. These details are essential for reproducing and interpreting the central measurements.
  3. The proposed functional form is not convincingly justified. Section 4.3 reports that alternative forms fit similarly, while Section 5.2 provides only a heuristic explanation for η/(1−η)\eta/(1-\eta) . This is insufficient to establish the specific equation as a quantitative causal law
  4. The experimental scope is limited relative to the generality claimed. Results mainly concern CIFAR/SVHN settings, and matched parameter counts still yield substantially different coefficients across architectures. Broader validation across dataset scales and model families is needed.
  5. The claim that memory is “fully invertible” (Line 312) is stronger than the recovery results support. After stripping, memorization remains at 0.11 and accuracy reaches 0.92 versus the FM0 reference of 0.96, demonstrating substantial recovery rather than complete inversion
Questions for authors

See weaknesses.

Reviewer_KbgYOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies neural networks under label noise training. They use a two-level persistence pipeline to separate the generalization channel from the memorization channel and quantify the effect of memorization on generalization (validation accuracy). They provide empirical results for this quantification formula and have experiments across various image classification datasets such as CIFAR-10, 100 and SVHN.

Critical strengths
  1. Some of the paper's observations are interesting. For example, Section 4.4 is of practical use: when there is a corrupted model which was trained on noisy labels, if we find a clean set of samples, we could restore generalization and revert memorization.
  2. The paper provides the experimental details, which help with the reproducibility of the results. The experiments are not limited to a single dataset, and are done over a few different empirical setups.
Critical weaknesses
  1. The paper is currently difficult to read. The overall flow and the ordering of the results obscure the meaning of various symbols, many of which are used without definition as early as the abstract. The abstract itself is unnecessarily long, burdened with specific details that lack a high-level interpretation. Furthermore, the results section begins with result C1, but its corresponding figure is relegated to the appendix, creating a disconnect between the written text and the visual presentation. Additionally, acronyms are not consistently defined. Overall, the manuscript is not sufficiently self-contained.
  2. The paper lacks any interpretation or practical implications of the introduced causal law between memorization and generalization. How can we interpret these results and make them useful in practice? What are the intuitions that we gain from this study? The study lacks providing any discussions around the results.
  3. Overall it is not clear what is the significance of the provided results? Why was a two-level persistence pipeline used to model the memorization and generalization behavior? The connection between the model and the actual learning trajectory during label noise training is not clear.
Questions for authors

How much is the coefficient C dependent on the label noise type as well as the dataset? There are various label noise schemas such as uniform, class-dependent, or feature-dependent, so it would be interesting to compare them in terms of the final causal law formula.

Reviewer_3LEKOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies label-noise memorization, its generalization cost and its connection with representation geometry. The authors identify an increase in cross class H_1 metric as memorization occurs and propose an empirical relationship between val-accuracy loss, memorization and clean/nose sample ratio.They further demonstrate substantial generalization recovery by continuing training on clean-only samples. Finally they propose a mechanism for the representation geometry change involving decision-boundary movement and displacement of clean representations.

Critical strengths

The empirical law describing the generalization cost of memorization is an interesting finding. The experiments also show that continuing training on clean-only samples can substantially reduce noisy-label memorization and recover validation accuracy.

Critical weaknesses
  1. It is difficult to distill actual contribution of the paper. The related-work discussion provides insufficient coverage of experimental studies on memorization, generalization, and representation geometry (see e.g., Stephenson et al., 2021; Maini et al., 2023). The paper should position its findings within this broader literature and clarify what its geometric and mechanistic claims add.

Cory Stephenson, suchismita padhy, Abhinav Ganesh, Yue Hui, Hanlin Tang, SueYeon Chung. On the geometry of generalization and memorization in deep neural networks. In: International Conference on Learning Representations, ICLR 2021. https://openreview.net/forum?id=V8jrrnwGbuc

Pratyush Maini, Michael Curtis Mozer, Hanie Sedghi, Zachary Chase Lipton, J Zico Kolter, Chiyuan Zhang. Can Neural Network Memorization Be Localized? Proceedings of the 40th International Conference on Machine Learning, PMLR 202:23536-23557, 2023. https://proceedings.mlr.press/v202/maini23a.html

  1. The methodology does not sufficiently define the core metrics or specify their computation. The manuscript does not precisely describe its within/cross-class topology measurements. Likewise, the representation-displacement measure (d_{\mathrm{clean}}) lacks an explicit definition and computation procedure. These omissions prevent a precise assessment of what the reported quantities establish and whether they support the paper’s geometric and mechanistic claims.

  2. The mechanistic interpretation is insufficiently described and supported. The authors identify the noisy to clean sample count-ratio with a “decision-boundary displacement ratio” without defining or measuring the latter. The feature-displacement measurement is also insufficiently specified, and its reported correlation with the accuracy-cost coefficient does not establish the proposed mechanism. Consequently, the section does not provide a precise, experimentally supported account connecting sample counts, representation movement, boundary changes, and accuracy loss.

Questions for authors

Next to the main criticism raised about, there are a few detailed questions:

  1. What is precisely meant by “memorization is addable”? (H_1^{\mathrm{within}}) is not shown, and preservation of the generalization channel is not demonstrated. How does the monotonic increase in (H_1^{\mathrm{cross}}) establish that memorization is independently addable, beyond showing that this statistic increases as the model fits corrupted labels?

  2. How did you establish that both CONT and overlay had finished converging? Does the reported accuracy gap persist if the overlay model is trained longer or with a different learning-rate schedule?

  3. Fig. 1a: Why is the SVHN full-overlay result not shown?

  4. Fig. 7: Does the decrease in memorization cost with increasing model size also hold within the other architecture families? Experiments at multiple sizes within each family would help distinguish the effects of model size from architecture.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.