Pith. sign in

REVIEW 2 major objections 4 minor 42 references

Random Label Prediction Heads for Studying Memorization in Deep Neural Networks

T0 review · 2 major / 4 minor · reviewed 2026-07-14 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Reducing memorization with random-label heads can either help or hurt generalization, depending on how densely the training data samples the true distribution.

desk verdict Clean operational probe for layer-wise memorization plus a regularizer that actually flips the sign of the generalization effect with sampling density; the Rademacher framing is loose but not load-bearing. read the letter →

arxiv 2607.11541 v1 pith:R2LAAYKL submitted 2026-07-13 cs.LG

Marlon Becker , Jonas Konrad , Luis Garcia Rodriguez , Benjamin Risse This is my paper · ORCID
classification cs.LG
keywords memorizationRademachercomplexitygeneralizationoverfittingdeepneuralnetworksregularizationrandomlabelslayer-wiseanalysis
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

Deep networks can memorize individual training samples so thoroughly that a simple extra head can recover fixed random labels from their intermediate features. The paper turns that fact into a practical measuring stick: attach a random-label prediction head (RLP-head) at any depth, read its accuracy as an empirical estimate of memorization (and of Rademacher complexity), and, if desired, reverse the loss so the network is punished for succeeding. The resulting regularizer cleanly suppresses memorization. Yet the effect on test accuracy is not uniform: it improves generalization on large, well-sampled data such as ImageNet and degrades it on smaller or long-tailed sets such as CIFAR-100. The authors therefore argue that memorization is not synonymous with harmful overfitting; its value depends on whether the data already cover the underlying distribution densely enough for shared features to be reliable.

What carries the argument

The random label prediction head (RLP-head): a shallow classifier attached in parallel to any chosen layer that is trained solely to predict fixed per-sample random labels from that layer’s activations. Its accuracy estimates memorization; a reverse-cross-entropy loss computed from its outputs regularizes the feature extractor against sample-specific features.

What would settle it

Train the identical architecture on a fully synthetic classification problem whose true decision boundary is known and densely sampled, apply RLP-regularization at increasing strengths, and check whether test error rises exactly when random-label accuracy falls to chance; if test error continues to fall while memorization is driven to zero, the sampling-density claim is falsified.

Watch

Extended reading notes

Core claim

Random-label prediction heads give a direct, layer-wise empirical measure of memorization capacity. When their success is actively suppressed by a reverse cross-entropy regularizer, memorization falls, but test performance rises on sufficiently sampled datasets and falls on undersampled ones. Memorization is therefore not equivalent to overfitting; it can be either detrimental or beneficial according to the sampling density of the training distribution.

Load-bearing premise

That the accuracy of a simple linear head at recovering fixed random labels from a layer’s activations faithfully measures how much sample-specific information that layer has memorized, rather than being limited by the head’s own capacity or by accidental structure in the random labels.

Editorial extensions

If this is right

  • Memorization can be monitored layer-by-layer during ordinary supervised training without altering the main task loss.
  • Explicit anti-memorization regularization can raise test accuracy and shrink the train–test gap on large, well-sampled vision datasets.
  • On undersampled or long-tailed data the same regularizer can destroy accuracy on rare classes that require sample-specific features.
  • Regularizing only the final layer shifts both class abstraction and residual memorization into earlier layers, sometimes improving early-exit accuracy.
  • Classical complexity bounds that treat lower Rademacher complexity as strictly better do not capture the dual role of memorization observed here.

Reading between the lines

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

  • The same RLP probe could be attached to intermediate activations of large language models to separate rare-document memorization from shared linguistic structure.
  • If sampling density decides whether memorization helps, deliberate densification of long-tail regions may be a more effective alternative to anti-memorization regularizers.
  • Layer-wise RLP accuracy curves offer a cheap diagnostic for deciding when a network is oversized relative to a given dataset.
  • Multi-head random partitions give stronger regularization, suggesting residual shared structure still leaks into the single-head metric and could be refined.
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

2 major / 4 minor

Summary. The paper introduces Random Label Prediction Heads (RLP-heads): auxiliary linear (or shallow) heads attached to intermediate activations that are trained to predict fixed random labels assigned to each training sample. Random-label accuracy is treated as an operational proxy for memorization capacity (motivated by Rademacher complexity). The same head yields a regularizer L_reg that penalizes confident random-label predictions and is applied only to the feature extractor. Extensive experiments on ViT-B/32–ImageNet and WideResNet-16-4–CIFAR-100 (plus ablations on head capacity, multi-head variants, frozen extractors, dataset fractions, label noise, ImageNet-LT, mixup, adversarial robustness and membership inference) show that RLP accuracy tracks known regularizers and capacity, that L_reg can drive it to chance, and that suppressing memorization improves test accuracy on sufficiently sampled data while degrading it on undersampled data. The authors hypothesize that the dual role is driven by sampling density and long-tail structure.

Significance. If the dual-role pattern holds, the work supplies both a practical, non-intrusive probe of layer-wise memorization and a controllable regularizer that can be used to test when memorization helps or hurts. The released code, the multi-control experimental design (dataset-fraction, noise, long-tail, de-duplication, multi-head, frozen re-probes), and the explicit separation of the metric from any theoretical generalization bound are genuine strengths. The result challenges the still-common identification of overfitting with memorization and offers a concrete experimental handle for studying long-tail effects (Feldman-style) without requiring leave-one-out retraining.

major comments (2)
  1. The dual-role claim (ImageNet improves, CIFAR-100 degrades under RLP-regularization) is the central empirical result, yet most main-text figures (Figs. 4–5, 7–9) report single-run curves without error bars or multi-seed statistics. Appendix A.17–A.18 give means±std only for the adversarial and membership-inference side experiments. Because the claimed effect sizes are modest (≈1.5 % on ImageNet, several points on CIFAR), at least three independent seeds for the key λ-sweeps on both datasets are needed to establish that the sign difference is robust rather than run-to-run variation.
  2. Section 3–4 present random-label accuracy as “an empirical estimate of Rademacher complexity.” The formal Rademacher definition (Eq. 1) involves a supremum over the hypothesis class and an expectation over random signs; the paper’s operational proxy is the accuracy of a single linear head trained by SGD on a fixed multi-class random labeling. While the proxy is later validated empirically (capacity sweeps, full-network random training in A.15, noisy-label correlation), the theoretical language should be softened to “operational proxy motivated by Rademacher complexity” so that readers do not treat the PAC bound of Theorem 1 as directly applicable to the reported numbers.
minor comments (4)
  1. Figure 2C and the layer-wise plots in Fig. 8 would benefit from a brief note on whether the RLP-heads share parameters or are independently initialized; the text is currently silent.
  2. Appendix A.12 reports Pearson r = 0.08 between RLP accuracy and Feldman scores; a short discussion of why the two measures diverge (generalization vs pure memorization) is already present but could be moved or cross-referenced more prominently in the main text.
  3. Typographical inconsistencies appear in the arXiv rendering (“RANDOMLABELPREDICTIONHEADS”, missing spaces after periods in several appendix headings). A light copy-edit pass would improve readability.
  4. The choice n = 100 000 for ImageNet (≈12 samples per random label) is justified in A.4–A.5, but a one-sentence reminder in the main experimental section would help readers who skip the appendix.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RLP accuracy is an operational proxy validated against independent controls; dual-role generalization effects are empirical observations, not forced by construction or self-citation.

full rationale

The paper's central chain is definitional only in the weak, non-circular sense that Rademacher complexity is classically the capacity to fit random labels (Sec. 3, Eq. 1), so measuring accuracy of an auxiliary linear head on fixed random labels is a natural empirical proxy (Sec. 4). This is not self-definitional circularity: the authors never claim a mathematical derivation that recovers a bound or uniqueness theorem from the metric itself; they treat random-label accuracy solely as a probe, then validate it against independent quantities (capacity sweeps A.9, full-network random-label training A.15 with Pearson r=0.93, noisy-label accuracy A.13–A.14, Feldman scores A.12, known regularizers Fig. 3). The regularizer L_reg (Eq. 4) is an explicit adversarial penalty on that probe; its effects on train/test accuracy are measured experimentally across datasets, fractions, noise levels, and long-tail classes (Figs. 4–7, A.2, A.19, A.21). No free parameter is fitted to a subset and then re-presented as a prediction; λ is swept and results reported for all values. No load-bearing self-citation, uniqueness import, or ansatz smuggling appears. The dual-role claim (memorization helps or hurts depending on sampling) is therefore an empirical pattern corroborated by multiple orthogonal ablations, not a result that reduces to its inputs by construction. Score 0 is the correct honest finding.

Assumptions & free parameters 3 free parameters · 3 assumptions · 1 invented entities

The central claims rest on standard PAC/Rademacher theory plus the operational definition of the RLP-head; free parameters are the usual training hyper-parameters plus the new regularization strength λ and the number of random labels n. No new physical entities are postulated.

free parameters (3)
  • regularization strength λ
    Tunable scalar that multiplies L_reg; chosen by grid search for each architecture/dataset pair; optimal values differ by orders of magnitude (10^4 on ImageNet, ~0.1 on CIFAR).
  • number of random labels n
    Chosen by hand (10k for CIFAR, 100k for ImageNet) to balance head capacity against compute; shown to affect measured accuracy (Appendix A.4).
  • RLP-head architecture (linear vs two-layer, hidden size)
    Default is a single linear layer; capacity ablations appear in A.6; results are qualitatively robust once capacity is sufficient.
assumptions (3)
  • domain assumption Empirical Rademacher complexity can be approximated by the accuracy of a linear head trained on fixed random labels while the feature extractor is trained on true labels.
    Stated in Section 3–4 as the theoretical motivation; validated empirically but not proved as a tight bound.
  • ad hoc to paper Only the RLP-head receives gradients from L_rnd; the feature extractor receives gradients from L_reg but not from the classification head of the RLP path.
    Architectural choice that isolates the memorization signal (Figure 1 and Section 4).
  • domain assumption Random labels are drawn once and held fixed across epochs.
    Standard for Rademacher-style estimates; sanity-checked by the shuffled-label ablation (A.2).
invented entities (1)
  • Random Label Prediction Head (RLP-head)
    purpose: Auxiliary head that predicts fixed random labels from intermediate activations, serving both as a memorization metric and as the source of the regularization signal.
    New architectural component introduced in Section 4; no independent existence outside the paper’s experimental setup.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Random Label Prediction Heads for Studying Memorization in Deep Neural Networks." pith.science (2026). https://pith.science/paper/R2LAAYKL

@misc{pith2026260711541,
  author       = {Pith},
  title        = {Pith review of: Random Label Prediction Heads for Studying Memorization in Deep Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R2LAAYKL}},
  note         = {Machine review of arXiv:2607.11541}
}
read the original abstract

We introduce a straightforward yet effective method to empirically study memorization in deep neural networks for classification tasks. Our approach augments each training sample with auxiliary random labels, which are then predicted by a random label prediction head (RLP-head). RLP-heads can be attached at arbitrary depths of a network, predicting random labels from the corresponding intermediate representation and thereby enabling analysis of how memorization capacity evolves across layers. By interpreting the RLP-head performance as an empirical estimate of Rademacher complexity, we obtain a direct measure of both sample-level memorization and model capacity. We leverage this random label accuracy metric to analyze generalization and overfitting in different models and datasets. Building on this approach, we further propose a novel regularization technique based on the output of the RLP-head, which demonstrably reduces memorization. Interestingly, our experiments reveal that reducing memorization can either improve or impair generalization, depending on the dataset and training setup. These findings challenge the traditional assumption that overfitting is equivalent to memorization and suggest new hypotheses to reconcile these seemingly contradictory results. The source code is available at https://github.com/MarlonBecker/RandomLabelHeads

Figures

Figures reproduced from arXiv: 2607.11541 by the authors.

Figure 1
Figure 1. An additional Random Label Prediction Head (RLP-head) is added after the feature extrac￾tor of the network. Only the RLP-head receives L rnd, the random label prediction loss, whereas the regularizing loss L reg is calculated on the RLP￾head but acts on the feature extractor only. In order to train the RLP-head we introduce an auxiliary cross-entropy loss on the random labels, L rnd, in addition to the standard clas… view at source ↗
Figure 2
Figure 2. ViT-B/32 on ImageNet. A: The proposed single fully-connected layer as RLP-head is sufficient to correctly predict approx. 70 % of the random labels, indicating that the feature extractor memorizes a substantial portion of the training set. B: Even after freezing the feature extractor, RLP-head attains low accuracy in the early epochs, confirming that the default RLP-head approach reliably tracks the evolution of mem… view at source ↗
Figure 3
Figure 3. WRN16-4 on CIFAR-100. The effect of common complexity regularizers can be measured [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: We find that RLP-regularization effectively suppresses memorization in both experimental [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 4
Figure 4. Figure 4: ViT-B/32 on ImageNet. Random label, train and test accuracy under RLP-regularization for [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: Schematic illustration of how memorization can be either detrimental or benign depending [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: ViT-B/32 on ImageNet. A+B: Random label and test accuracy when training on reduced dataset fractions (DF) of ImageNet. Although smaller training sets lead to stronger memorization (higher random label accuracy), suppressing memorization on them does not improve test ac…
Figure 8
Figure 8. Figure 8: ViT-B/32 on ImageNet. A: Random label accuracy of RLP-heads at different layers when only the final (12th) layer is used for RLP-regularization. Memorization shifts toward earlier layers. B+C: RLP-regularization is calculated based on RLP-heads attached to all 12 trans…
Figure 9
Figure 9. Figure 9: ViT-B/32 on ImageNet. Similar to the RLP-heads, we attach additional classification heads [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references

  1. [1]

    Conference on Computer Vision and Pattern Recognition , year =

    Christian Szegedy and Vincent Vanhoucke and Sergey Ioffe and Jonathon Shlens and Zbigniew Wojna , title =. Conference on Computer Vision and Pattern Recognition , year =

  2. [2]

    International Conference on Learning Representations , year=

    Decoupled Weight Decay Regularization , author=. International Conference on Learning Representations , year=

  3. [3]

    ImageNet:

    Jia Deng and Wei Dong and Richard Socher and Li. ImageNet:. Conference on Computer Vision and Pattern Recognition , year =

  4. [4]

    International Conference on Learning Representations , year =

    Alexey Dosovitskiy and Lucas Beyer and Alexander Kolesnikov and Dirk Weissenborn and Xiaohua Zhai and Thomas Unterthiner and Mostafa Dehghani and Matthias Minderer and Georg Heigold and Sylvain Gelly and Jakob Uszkoreit and Neil Houlsby , title =. International Conference on Learning Representations , year =

  5. [5]

    Entropy and mutual information in models of deep neural networks , journal =

    Marylou Gabri. Entropy and mutual information in models of deep neural networks , journal =

  6. [6]

    British Machine Vision Conference , year =

    Sergey Zagoruyko and Nikos Komodakis , title =. British Machine Vision Conference , year =

  7. [7]

    Learning Multiple Layers of Features from Tiny Images , year =

    Krizhevsky, Alex and Hinton, Geoffrey , institution =. Learning Multiple Layers of Features from Tiny Images , year =

  8. [8]

    ACM Comput

    Mehrabi, Ninareh and Morstatter, Fred and Saxena, Nripsuta and Lerman, Kristina and Galstyan, Aram , title =. ACM Comput. Surv. , pages =. 2021 , volume =

    2021

Show all 42 references
  1. [9]

    Journal of Machine Learning Research , year =

    Nitish Srivastava and Geoffrey Hinton and Alex Krizhevsky and Ilya Sutskever and Ruslan Salakhutdinov , title =. Journal of Machine Learning Research , year =

  2. [10]

    Neural Networks Learning and Memorization with (almost) no Over-Parameterization , year =

    Daniely, Amit , journal =. Neural Networks Learning and Memorization with (almost) no Over-Parameterization , year =

  3. [11]

    Foundations of machine learning

    Mohri, Mehryar and Rostamizadeh, Afshin and Talwalkar, Ameet. Foundations of machine learning

  4. [12]

    Zhang, Chiyuan and Bengio, Samy and Hardt, Moritz and Recht, Benjamin and Vinyals, Oriol , title =. Commun. ACM , pages =. 2021 , volume =

    2021

  5. [13]

    International Conference on Learning Representations , year=

    Identity Crisis: Memorization and Generalization Under Extreme Overparameterization , author=. International Conference on Learning Representations , year=

  6. [14]

    What Neural Networks Memorize and Why: Discovering the Long Tail via Influence Estimation , year =

    Feldman, Vitaly and Zhang, Chiyuan , journal =. What Neural Networks Memorize and Why: Discovering the Long Tail via Influence Estimation , year =

  7. [15]

    International Conference on Machine Learning , author =

    A. International Conference on Machine Learning , author =

  8. [16]

    2021 , volume =

    Preetum Nakkiran and Gal Kaplun and Yamini Bansal and Tristan Yang and Boaz Barak and Ilya Sutskever , title =. 2021 , volume =

    2021

  9. [17]

    Mitigating Unwanted Biases with Adversarial Learning , booktitle =

    Zhang, Brian Hu and Lemoine, Blake and Mitchell, Margaret , year =. Mitigating Unwanted Biases with Adversarial Learning , booktitle =

  10. [18]

    Wang, Zeyu and Qinami, Klint and Karakozis, Ioannis Christos and Genova, Kyle and Nair, Prem and Hata, Kenji and Russakovsky, Olga , year =. Towards

  11. [19]

    2019 , volume =

    Xue Ying , title =. 2019 , volume =

    2019

  12. [20]

    The secret sharer: evaluating and testing unintended memorization in neural networks , year =

    Carlini, Nicholas and Liu, Chang and Erlingsson, \'. The secret sharer: evaluating and testing unintended memorization in neural networks , year =. USENIX Conference on Security Symposium , pages =

  13. [21]

    Small ReLU networks are powerful memorizers: a tight analysis of memorization capacity , year =

    Yun, Chulhee and Sra, Suvrit and Jadbabaie, Ali , journal =. Small ReLU networks are powerful memorizers: a tight analysis of memorization capacity , year =

  14. [22]

    Image fairness in deep learning: problems, models, and challenges , journal =

    Tian, Huan and Zhu, Tianqing and Liu, Wei and Zhou, Wanlei , year =. Image fairness in deep learning: problems, models, and challenges , journal =

  15. [23]

    Memorization Without Overfitting: Analyzing the Training Dynamics of Large Language Models , year =

    Tirumala, Kushal and Markosyan, Aram and Zettlemoyer, Luke and Aghajanyan, Armen , journal =. Memorization Without Overfitting: Analyzing the Training Dynamics of Large Language Models , year =

  16. [24]

    Extracting Training Data from Large Language Models , booktitle =

    Nicholas Carlini and Florian Tram. Extracting Training Data from Large Language Models , booktitle =. 2021 , pages =

    2021

  17. [25]

    Proceedings of the 52nd Annual ACM SIGACT Symposium on Theory of Computing , volume =

    Vitaly Feldman , title =. Proceedings of the 52nd Annual ACM SIGACT Symposium on Theory of Computing , volume =

  18. [26]

    Mozer and Hanie Sedghi and Zachary C

    Pratyush Maini and Michael C. Mozer and Hanie Sedghi and Zachary C. Lipton and J. Zico Kolter and Chiyuan Zhang , year=. Can Neural Network Memorization Be Localized? , journal=

  19. [27]

    Deep Learning , author=

  20. [28]

    2024 , journal =

    The Pitfalls of Memorization: When Memorization Hurts Generalization , author=. 2024 , journal =

    2024

  21. [29]

    2021 , journal =

    Deep Learning Through the Lens of Example Difficulty , author=. 2021 , journal =

    2021

  22. [30]

    A Simple Weight Decay Can Improve Generalization , year =

    Krogh, Anders and Hertz, John , journal =. A Simple Weight Decay Can Improve Generalization , year =

  23. [31]

    Journal of Imaging , VOLUME =

    Barz, Björn and Denzler, Joachim , TITLE =. Journal of Imaging , VOLUME =. 2020 , NUMBER =

    2020

  24. [32]

    International Conference on Learning Representations , year=

    Towards Deep Learning Models Resistant to Adversarial Attacks , author=. International Conference on Learning Representations , year=

  25. [33]

    International Conference on Learning Representations , year=

    Explaining and Harnessing Adversarial Examples , author=. International Conference on Learning Representations , year=

  26. [34]

    Proceedings of the 37th International Conference on Machine Learning , year=

    Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks , author=. Proceedings of the 37th International Conference on Machine Learning , year=

  27. [35]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , year=

    Large-scale long-tailed recognition in an open world , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , year=

  28. [36]

    International Conference on Learning Representations , year=

    mixup: Beyond Empirical Risk Minimization , author=. International Conference on Learning Representations , year=

  29. [37]

    Early-Learning Regularization Prevents Memorization of Noisy Labels , year =

    Liu, Sheng and Niles-Weed, Jonathan and Razavian, Narges and Fernandez-Granda, Carlos , journal =. Early-Learning Regularization Prevents Memorization of Noisy Labels , year =

  30. [38]

    2016 , journal =

    Text Understanding from Scratch , author=. 2016 , journal =

    2016

  31. [39]

    RoBERTa:

    Yinhan Liu and Myle Ott and Naman Goyal and Jingfei Du and Mandar Joshi and Danqi Chen and Omer Levy and Mike Lewis and Luke Zettlemoyer and Veselin Stoyanov , year =. RoBERTa:

  32. [40]

    Scaling Learning Algorithms Towards

    Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

  33. [41]

    and Osindero, Simon and Teh, Yee Whye , journal =

    Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

  34. [42]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    2016

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.