Do We Train on Test Data? Purging CIFAR of Near-Duplicates

The CIFAR-10 and CIFAR-100 datasets are two of the most heavily benchmarked datasets in computer vision and are often used to evaluate novel methods and model architectures in the field of deep learning. However, we find that 3.3% and 10% of the images from the test sets of these datasets have duplicates in the training set. These duplicates are easily recognizable by memorization and may, hence, bias the comparison of image recognition techniques regarding their generalization capability. To eliminate this bias, we provide the "fair CIFAR" (ciFAIR) dataset, where we replaced all duplicates in the test sets with new images sampled from the same domain. The training set remains unchanged, in order not to invalidate pre-trained models. We then re-evaluate the classification performance of various popular state-of-the-art CNN architectures on these new test sets to investigate whether recent research has overfitted to memorizing data instead of learning abstract concepts. We find a significant drop in classification accuracy of between 9% and 14% relative to the original performance on the duplicate-free test set. We make both the ciFAIR dataset and pre-trained models publicly available and furthermore maintain a leaderboard for tracking the state of the art.