CoChat

Understanding deep learning (still) requires rethinking generalization

Open access
Zhang, Chiyuan, Bengio, Samy, Hardt, Moritz, Recht, Benjamin, Vinyals, Oriol · 2021 · Communications of the ACM

The DOI matches its Crossref record; 45 of the 47 references it cites were found in a registry or an index; none of the 12 sources its claims cite is checked for support yet.

Abstract

Despite their massive size, successful deep artificial neural networks can exhibit a remarkably small gap between training and test performance. Conventional wisdom attributes small generalization error either to properties of the model family or to the regularization techniques used during training. Through extensive systematic experiments, we show how these traditional approaches fail to explain why large neural networks generalize well in practice. Specifically, our experiments establish that state-of-the-art convolutional networks for image classification trained with stochastic gradient methods easily fit a random labeling of the training data. This phenomenon is qualitatively unaffected by explicit regularization and occurs even if we replace the true images by completely unstructured random noise. We corroborate these experimental findings with a theoretical construction showing that simple depth two neural networks already have perfect finite sample expressivity as soon as the number of parameters exceeds the number of data points as it usually does in practice. We interpret our experimental findings by comparison with traditional models. We supplement this republication with a new section at the end summarizing recent progresses in the field since the original version of this paper.

Overview

CoChat Summary

TL;DR

Deep neural networks can fit random labels and random pixels with zero training error, so traditional complexity and regularization explanations do not account for their generalization. The paper argues that explicit regularization is neither necessary nor sufficient, and offers a finite-sample expressivity construction plus linear-model implicit-regularization analysis.

Key findings

  • On CIFAR10, Inception, AlexNet, and MLPs reach 100% (or 99.82%/99.34%) training accuracy on random labels while test accuracy is near chance: Inception 9.78%, Inception w/o BatchNorm 10.12%, AlexNet 9.86%, MLP 3×512 10.48%, MLP 1×512 10.61%.
  • On ImageNet, Inception reaches 95.20% top-1 accuracy on 1.2 million random labels from 1000 categories without hyperparameter tuning; with explicit regularizers on it still reaches ∼90% top-1.
  • Explicit regularization may improve generalization but is neither necessary nor sufficient: on CIFAR10 all models still generalize well with regularizers off; on ImageNet top-1 accuracy without regularization is 59.80% vs 0.1% random guessing, and with data augmentation only Inception reaches 72.95% top-1.
  • Theoretically, a two-layer ReLU network with p = 2n + d parameters can represent any labeling of any sample of size n in d dimensions; depth-k networks can use O(n/k) parameters per layer. In linear models, SGD converges to a minimum-norm solution, but minimum norm is not predictive: on MNIST the minimum norm with no preprocessing is ∼220 and with wavelet preprocessing ∼390, yet test error drops by a factor of 2.

Methods

  • Randomization tests on CIFAR10 and ImageNet ILSVRC 2012: train Inception V3 on ImageNet and smaller Inception, AlexNet, and MLPs on CIFAR10 under true labels, partially corrupted labels, random labels, shuffled pixels, random pixels, and Gaussian pixels, toggling data augmentation, weight decay, and dropout.
  • Compare training/test accuracy and convergence time; supplement with a finite-sample expressivity theorem for two-layer ReLU networks and with linear/kernel experiments on MNIST and CIFAR10 (Gaussian kernels, random convolutional filters).

Limitations

  • Stated: the paper offers no new theory of generalization; it says a full mathematical characterization remains challenging and that it cannot fully survey recent progress. It also states that minimum-norm intuition is only a small piece of the generalization story and that it does not explain why certain architectures generalize better.
  • Evident: the empirical scope is limited to image classification benchmarks (CIFAR10, ImageNet) and a small set of architectures/regularizers, so the findings may not extend to other domains or model families. The theoretical construction uses a depth-2 network with large width, and the paper notes depth-k alternatives with O(n/k) parameters per layer.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3446776

Cites

45 of 45 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

0 Valid · 7 Not valid · 12 Not checked · 85 Own results and interpretations

70 more claims

Create a free account to see all 104 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 17
  • Probably not a substantial claim. 15
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Show all 45 references

Research works printed but not linked: 8

  1. Cohen, N., Shashua, A. Convolutional rectifier networks as generalized tensor decompositions. In International Conference on Machine Learning. M.F. Balcan and K.Q. Weinberger, eds. 2016, 955--963.
  2. Cybenko, G. Approximation by superposition of sigmoidal functions. Math. Control Signal. Syst. 4, 2 (1989), 303--314.
  3. Golowich, N., Rakhlin, A., Shamir, O. Size-independent sample complexity of neural networks. In Conference On Learning Theory. P.R. Sébastien Bubeck, V. Perchet, eds., 2018, 297--299.
  4. Ma, S., Bassily, R., Belkin, M. The power of interpolation: Understanding the effectiveness of sgd in modern over-parametrized learning. In International Conference on Machine Learning. J. Dy and A. Krause, eds. 2018, 3325--3334.
  5. Morcos, A., Raghu, M., Bengio, S. Insights on representational similarity in neural networks with canonical correlation. Adv. Neural Inform. Process. Syst.. 2018, 5727--5736.
  6. Neyshabur, B., Tomioka, R., Srebro, N. Norm-based capacity control in neural networks. In Conference on Learning Theory. S.K. Peter Grünwald and E. Hazan, eds. 2015, 1376--1401.
  7. Smith, S.L., Le, Q.V. A bayesian perspective on generalization and stochastic gradient descent. In International Conference on Learning Representations, 2018.
  8. Zhou, W., Veitch, V., Austern, M., Adams, R.P., Orbanz, P. Non-vacuous generalization bounds at the ImageNet scale: A PAC-Bayesian compression approach. In International Conference on Learning Representations, 2019.

Cited by

Items 2,101–2,200 of 2,667

Metadata

Authors
Zhang, C., Bengio, S., Hardt, M., Recht, B., & Vinyals, O.
Affiliations
  • Google (United States) (US)
  • Google Brain, Mountain View, CA
  • University of California, Berkeley (US)
  • Google DeepMind (United Kingdom) (GB)
Venue
Communications of the ACM
Published
2021-02-22
Volume
64
Issue
3
Pages
107-115
ISSN
  • 0001-0782 (ISSN-L)
  • 1557-7317
OpenAlex
W2566079294
Cited by
2,400
References
47
Citations by year
2026: 295 · 2025: 558 · 2024: 579 · 2023: 500 · 2022: 316 · 2021: 108
FWCI
132.37
Type
journal article
Language
en
Open access
Locations
Communications of the ACM (publishedVersion, cc-by-nc-sa)
Funders
Google
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Stochastic Gradient Optimization Techniques
  • Physical Sciences › Computer Science › Artificial Intelligence › Neural Networks and Applications
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
Keywords
generalization, deep neural network, overfitting, overparameterization, expressivity, stochastic gradient descent, regularization, image classification, convolutional neural networks

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-08, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-08 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.