Trimmed Robust Loss Function for Training Deep Neural Networks with Label Noise

摘要

Deep neural networks obtain nowadays outstanding results on many vision, speech recognition and natural language processing-related tasks. Such deep structures need to be trained on very large datasets, what makes annotating the data for supervised learning, particularly difficult and time-consuming task. In the supervised datasets label noise may occur, which makes the whole training process less reliable. In this paper we present a novel robust loss function based on categorical cross-entropy. We demonstrate its robustness for several amounts of noisy labels, on popular MNIST and CIFAR-10 datasets.