Correlated Input-Dependent Label Noise in Large-Scale Image Classification

Large scale image classification datasets often contain noisy labels. We take\na principled probabilistic approach to modelling input-dependent, also known as\nheteroscedastic, label noise in these datasets. We place a multivariate Normal\ndistributed latent variable on the final hidden layer of a neural network\nclassifier. The covariance matrix of this latent variable, models the aleatoric\nuncertainty due to label noise. We demonstrate that the learned covariance\nstructure captures known sources of label noise between semantically similar\nand co-occurring classes. Compared to standard neural network training and\nother baselines, we show significantly improved accuracy on Imagenet ILSVRC\n2012 79.3% (+2.6%), Imagenet-21k 47.0% (+1.1%) and JFT 64.7% (+1.6%). We set a\nnew state-of-the-art result on WebVision 1.0 with 76.6% top-1 accuracy. These\ndatasets range from over 1M to over 300M training examples and from 1k classes\nto more than 21k classes. Our method is simple to use, and we provide an\nimplementation that is a drop-in replacement for the final fully-connected\nlayer in a deep classifier.\n