Do CIFAR-10 Classifiers Generalize to CIFAR-10?

By B. Recht, Rebecca Roelofs, Ludwig Schmidt, Vaishaal Shankar. 2018

470 citations recorded by the source corpus.

Machine learning is currently dominated by largely experimental work focused on improvements in a few key tasks. However, the impressive accuracy numbers of the best performing models are questionable because the same test sets have been used to select these models for multiple years now. To understand the danger of overfitting, we measure the accuracy of CIFAR-10 classifiers by creating a new test set of truly unseen images. Although we ensure that the new test set is as close to the original data distribution as possible, we find a large drop in accuracy (4% to 10%) for a broad range of deep learning models. Yet more recent models with higher original accuracy show a smaller drop and better overall performance, indicating that this drop is likely not due to overfitting based on adaptivity. Instead, we view our results as evidence that current accuracy numbers are brittle and susceptible to even minute natural variations in the data distribution.

Subfields: Computer Vision and Pattern Recognition, Evaluation & Benchmarks

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Evaluation & Benchmarks

Citations in the corpus

Adversarial NLI: A New Benchmark for Natural Language Understanding (2019, background): “Machine learning also has a tendency to overﬁt on static targets, even if that does not happen deliberately (Recht et al., 2018).”
Deduplicating Training Data Mitigates Privacy Risks in Language Models (2022, background): “Consequently, deduplication is still necessary even when training with DP. Duplication Beyond Text Data Our work focuses on nat-ural language, but datasets in domains such as images and source code also contain duplicate samples (Recht et al., 2018; Ziegler, 2021).”
A Large-scale Study of Representation Learning with the Visual Task Adaptation Benchmark (2019, background): “Fortunately, despite numerous test re-evaluations on popular ML benchmarks, such as ImageNet, progress seems to transfer to new data (Recht et al., 2018; 2019; Kornblith et al., 2019).”
Benchmarking Neural Network Robustness to Common Corruptions and Perturbations (2019, background): “The human vision system is robust in ways that existing computer vision systems are not (Recht et al., 2018; Azulay & Weiss, 2018).”
What Can ResNet Learn Efficiently, Going Beyond Kernels? (2019, methodology): “For example, ResNet achieves 96% test accuracy on the CIFAR-10 data set, but NTKs achieve 77% [6] and random feature kernels achieve 85% [29].”
AutoAugment: Learning Augmentation Policies from Data (2018, methodology): “We also evaluate our best model trained with AutoAugment on a recently prepared CIFAR-10 test set [26].”
nocaps: novel object captioning at scale (2019, background): “However, recent work has observed significant performance degradation when transferring models across datasets even when the new target dataset is an exact recreation of the old dataset [33].”
Differentially Private Learning Needs Better Features (or Much More Data) (2020, background): “Indeed, we had originally considered a differentially-private variant of the random-feature CIFAR-10 classifier proposed in (Recht et al., 2018), but found the model’s high dimensionality (over 10 million features) to be detrimental to private learning.”
sharpDARTS: Faster and More Accurate Differentiable Architecture Search (2019, methodology): “We also show a statistically significant[22] improvement over ShakeShake64d[22, 7] which is the best available CIFAR-10.”
Deep Weighted Averaging Classifiers (2018, background): “Finally, careful evaluation and an explosion of work on adversarial examples have demonstrated that many deep models are more brittle than test-set accuracies would suggest [12, 15, 37, 41].”
Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: A cross-sectional study (2018, background): “that “current accuracy numbers are brittle and susceptible to even minute natural variations in the data distribution” [8].”
Unlabeled Data Improves Adversarial Robustness (2019, methodology): “Following [31], we define a near-duplicate as an image with `2 distance below 2000/255 to any of the CIFAR-10 test images.”
Identifying Statistical Bias in Dataset Replication (2020, methodology): “Following the protocol of Deng et al. (2009), Recht et al. (2019b) collected a pool of candidate image-label pairs, and estimated their selection frequencies via MTurk, along with a subset of the v1 validation set.”
Improving Robustness Without Sacrificing Accuracy with Patch Gaussian Augmentation (2019, background): “For example, neural networks are sensitive to small translations and changes in scale [2], blurring and additive noise [10], small objects placed in images [41], and even different images from a similar distribution of the training set [40, 39].”
Hidden stratification causes clinically meaningful failures in machine learning for medical imaging (2019, background): “Problems similar to hidden stratification have been observed or postulated in many domains, including traditional computer vision [30], fine-grained image recognition [39], genomics [7], and epidemiology (often termed “spectrum effects”) [23].”