Leveraging inductive bias of neural networks for learning without explicit human annotations

Classification problems today are typically solved by first collecting examples along with candidate labels, second obtaining clean labels from workers, and third training a large, overparameterized deep neural network on the clean examples. The second, labeling step is often the most expensive one as it requires manually going through all examples. In this paper we skip the labeling step entirely and propose to directly train the deep neural network on the noisy candidate labels and early stop the training to avoid overfitting. With this procedure we exploit an intriguing property of large overparameterized neural networks: While they are capable of perfectly fitting the noisy data, gradient descent fits clean labels much faster than the noisy ones, thus early stopping resembles training on the clean labels. Our results show that early stopping the training of standard deep networks such as ResNet-18 on part of the Tiny Images dataset, which does not involve any human labeled data, and of which only about half of the labels are correct, gives a significantly higher test performance than when trained on the clean CIFAR-10 training dataset, which is a labeled version of the Tiny Images dataset, for the same classification problem. In addition, our results show that the noise generated through the label collection process is not nearly as adversarial for learning as the noise generated by randomly flipping labels, which is the noise most prevalent in works demonstrating noise robustness of neural networks.