Understanding deep learning requires rethinking generalization
By Chiyuan Zhang, Samy Bengio, Moritz Hardt, B. Recht, Oriol Vinyals. 2016
5,191 citations recorded by the source corpus.
Despite their massive size, successful deep artificial neural networks can exhibit a remarkably small difference 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.
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
- Yoshua Bengio
- Mohit Bansal
- Percy Liang
- Quoc V. Le
- Sergey Levine
- Jianfeng Gao
- Ruslan Salakhutdinov
- Luke S. Zettlemoyer
In Scaling & Learning Theory
Citations in the corpus
- Train Large, Then Compress: Rethinking Model Size for Efficient Training and Inference of Transformers (2020, background): “For instance, it studies forms of implicit regularization (Zhang et al., 2017; Belkin et al., 2018), characterizes the expressivity of deep models (Raghu et al.”
- Deep Learning Scaling is Predictable, Empirically (2017, neutral): “While these theoretical and empirical results offer significant insight about required model sizing, recent work has noted the need for more practical guidance (Zhang et al. (2017); Arpit et al. (2017); Smith and Le (2017); Kawaguchi et al. (2017)).”
- Scaling description of generalization with number of parameters in deep learning (2019, background): “Surprisingly from the point of view of traditional statistical learning theory [5] DNNs generalize very well in practice, even without an explicit regularization.”
- Gradient Descent Happens in a Tiny Subspace (2018, methodology): “…activation function (e.g. softplus instead of ReLU), models trained using different optimization algorithms (SGD and Adam), models trained using different batch sizes and learning rates, models trained on data with random labels (as was considered by Zhang et al. (2016)), and a regression task.”
- Neural Tangent Kernel: Convergence and Generalization in Neural Networks (2018, background): “It seems paradoxical that a reasonably large neural network can fit random labels, while still obtaining good test accuracy when trained on real data [19].”
- High-dimensional dynamics of generalization error in neural networks (2017, background): “An empirical approach to test for memorization is to see if a deep neural network can fit a training set with randomly scrambled labels rather than the true labels (Zhang et al., 2017).”
- Quantifying Memorization Across Neural Language Models (2022, methodology): “In particular, our analysis on model scale is informed by preliminary experiments in (Zhang et al., 2017; Carlini et al., 2020), our data duplication experiments follow in the line of Lee et al.”
- GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding (2020, background): “Across different model families, it has also been observed that bigger models with larger model capacities not only fit the training data better but also generalize better on test time (Zhang et al., 2017; Neyshabur et al., 2017; Huang et al., 2019).”
- Don't Decay the Learning Rate, Increase the Batch Size (2017, background): “However while SGD finds minima that generalize well (Zhang et al., 2016; Wilson et al., 2017), each parameter update only takes a small step towards the objective.”
- Extracting Training Data from Large Language Models (2020, background): “Such privacy leakage is typically associated with overfitting [69]—when a model’s training error is significantly lower than its test error—because overfitting often indicates that a model has memorized examples from its training set.”
- ShapeWorld - A new test methodology for multimodal language understanding (2017, background): “Investigations for image recognition (Szegedy et al., 2014; Nguyen et al., 2015; Zhang et al., 2017) have shown surprising behavior very different from what would be expected of their “surpassing human-level performance” (He et al.”
- mixup: Beyond Empirical Risk Minimization (2017, methodology): “Following Zhang et al. (2017), we evaluate the robustness of ERM and mixup models against randomly corrupted labels.”
- Sharpness-Aware Minimization for Efficiently Improving Generalization (2020, background): “Indeed, many modern neural networks can easily memorize the training data and have the capacity to readily overfit (Zhang et al., 2016).”
- Multi-task Self-Supervised Visual Learning (2017, background): “Unlabeled images and video are available in practically unlimited quantities, and the most prominent present image models—neural networks—are data starved, easily memorizing even random labels for large image collections [45].”
- Scalable agent alignment via reward modeling: a research direction (2018, methodology): “Since we cannot train the reward model and the agent on all possible outcomes, we need it to generalize appropriately from the given data (Zhang et al., 2017, 2018a).”
- Stochastic Polyak Step-size for SGD: An Adaptive Learning Rate for Fast Convergence (2020, background): “Modern machine learning models such as non-parametric regression or over-parametrized deep neural networks are highly expressive and can fit or interpolate the training dataset completely (Zhang et al., 2017; Ma et al., 2018).”
- Fast and Furious Convergence: Stochastic Second Order Methods under Interpolation (2019, background): “, 1998), and over-parameterized deep neural networks (Zhang et al., 2017).”
- Painless Stochastic Gradient: Interpolation, Line-Search, and Convergence Rates (2019, methodology): “The interpolation condition is satisfied for models such as non-parametric regression [7, 45], over-parametrized deep neural networks [88], boosting [69], and for linear classifiers on separable data.”
- Implicit Regularization of Discrete Gradient Dynamics in Deep Linear Neural Networks (2019, background): “The implicit regularization provided by the choice of the optimization algorithm has recently become an active area of research in machine learning, putting lot of interest on the behavior of gradient descent on deep over-parametrized models [Neyshabur et al., 2015b, 2017, Zhang et al., 2017].”
- A Closer Look at Memorization in Deep Networks (2017, background): “2017 [1] ● Shows: DNNs can fit random labels .”