Dimensionality-Driven learning with noisy labels

© The Author(s) 2018. Datasets with significant proportions of noisy (incorrect) class labels present challenges for training accurate Deep Neural Networks (DNNs). We propose a new perspective for understanding DNN generalization for such datasets, by investigating the dimensionality of the deep represen-tation subspace of training samples. We show that from a dimensionality perspective, DNNs ex-hibit quite distinctive learning styles when trained with clean labels versus when trained with a pro-portion of noisy labels. Based on this finding, we develop a new dimensionality-driven learning strategy, which monitors the dimensionality of subspaces during training and adapts the loss function accordingly. We empirically demonstrate that our approach is highly tolerant to significant proportions of noisy labels, and can effectively learn low-dimensional local subspaces that capture the data distribution.