Understanding deep learning (still) requires rethinking generalization
By Chiyuan Zhang, Samy Bengio, Moritz Hardt, B. Recht, Oriol Vinyals. 2021
2,784 citations recorded by the source corpus.
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.
Subfields: Artificial Intelligence, Scaling & Learning Theory
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Artificial Intelligence
- Jianfeng Gao
- Percy Liang
- Mohit Bansal
- Philip S. Yu
- Yoshua Bengio
- Dacheng Tao
- Luke S. Zettlemoyer
- Yonatan Belinkov
In Scaling & Learning Theory
- Tal Linzen
- R. Thomas McCoy
- Tongliang Liu
- Simon Lacoste-Julien
- Prateek Mittal
- Philip S. Yu
- Shuang Song
- James Y. Zou
Citations in the corpus
- Dynamics of SGD with Stochastic Polyak Stepsizes: Truly Adaptive Variants and Convergence to Exact Solution (2022, background): “In the non-regularized overparametrized setting (e.g. neural networks), f S k is often zero for every subset S [34].”
- Beyond Not-Forgetting: Continual Learning with Backward Knowledge Transfer (2022, background): “This definition of sufficient projection follows the same line of the trust region defined in [17], which implies that task t and j may have sufficient common bases between their input subspaces and hence are strongly correlated, because the gradient lies in the span of the input [33].”
- Online Continual Learning on a Contaminated Data Stream with Blurry Task Boundaries (2022, background): “However, on the contaminated data stream, diversity-based memory construction may promote to include many examples with corrupted labels in the episodic memory, leading to poor performance thanks to very high capacity of deep models to overly fit all the falsely labeled examples [40].”
- TRGP: Trust Region Gradient Projection for Continual Learning (2022, background): “Based on the fact that stochastic gradient descent updates lie in the subspace spanned by the data input (Zhang et al., 2021; Saha et al., 2021), it follows that the gradient Therefore, the projection of ∇ W l L 2 onto the orthogonal direction to S l 1 is 0, which means that the model W l 1 will…”
- SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training (2025, neutral): “…RL are both widely used for foundation model training (OpenAI, 2023b; Google, 2023; Jaech et al., 2024; DeepSeekAI et al., 2025), their distinct effects on generalization (Bousquet & Elisseeff, 2000; Zhang et al., 2021) remain unclear, making it challenging to build reliable and robust AI systems.”
- On Measuring Excess Capacity in Neural Networks (2022, background): “The observation that (over-parametrized) models can easily fit—i.e., reach zero training error—to randomly permuted training labels [45, 46] but, when trained on unpermuted labels, yield good generalization performance, has fueled much of the progress in this area.”
- Learning with Noisy Labels Revisited: A Study Using Real-World Human Annotations (2021, methodology): “There are some techniques for synthesizing the instance-dependent label noise, such as the polynomial margin diminishing label noise [Zhang et al., 2021b] where instances near decision boundary are easier to be mislabeled, the part-dependent label noise [Xia et al., 2020b] where different parts of…”
- Towards Unbounded Machine Unlearning (2023, background): “…influence of a subset of the training set from the weights of a trained model is hard, since deep models memorize information about specific instances [Zhang et al., 2020, 2021, Arpit et al., 2017] and their highly non-convex nature makes it difficult to trace the effect of each example on the…”
- Membership Inference Attacks on Machine Learning: A Survey (2021, background): “Machine learning (ML) models such as DNNs are often overparameterized, which means that they have sufficient capacity to memorize information about their training dataset [16, 144, 183, 234].”
- Privacy and Robustness in Federated Learning: Attacks and Defenses (2020, background): “Hence, GAN attack is not targeted to reconstruct the exact training inputs, but only the class representatives.”
- Bag of Tricks for Training Data Extraction from Language Models (2023, background): “The presence of a high degree of confidence in memorized data is one of the defining features of the phenomenon known as the’memorization effect’ (Goodfellow et al., 2016; Zhang et al., 2021; Arpit et al., 2017).”
- Poisoning Web-Scale Training Datasets is Practical (2023, background): “This quantity-over-quality tradeoff has so far been deemed acceptable, both because modern neural networks are extremely resilient to large amounts of label noise [79], [114], and because training on noisy data can even improve model utility on out-of-distribution data [74], [75].”
- Aging with GRACE: Lifelong Model Editing with Discrete Key-Value Adaptors (2022, neutral): “Beyond unrealistically assessing performance, it is well-established that learning random versus natural labels is vastly different [49].”
- Shifting machine learning for healthcare from development to deployment and from models to data (2022, background): “In certain instances, data leakage can occur, such as when ML models ‘memorize’ datasets 80–82 and when access to model parameters and updates can be used to infer the original dataset 83 .”
- Erasing Concepts from Diffusion Models (2023, background): “While the traditional goal of machine learning is to generalize without memo-rization, large models are capable of exact memorization if specifically trained to do so [48], and unintentional memo-rization has also been observed in large-scale settings [6, 5], including diffusion models [45].”
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (2023, background): “Over-parameterized neural networks have the capacity to fit almost any data points, including randomly labeled training data (Feldman & Zhang, 2020; Zhang et al., 2021).”
- InfLoRA: Interference-Free Low-Rank Adaptation for Continual Learning (2024, background): “GPM and DualGPM are established on the fact that the gradient updates lie in the span of input data points [8].”