TRAINING DEEP NEURAL-NETWORKS USING A NOISE ADAPTATION LAYER ICLR 2017 conference submission
Learning MNIST when almost half the labels are permuted in a fixed way. For example, when the task of labeling is split between two people that don’t agree.
Follow mnist-simple notebook for an example of how to implement the Simple noise adaption layer in the paper with a single customized Keras layer. Follow 161103-run-plot, 161202-run-plot-cifar100 and 161230-run-plot-cifar100-sparse notebooks for how to reproduce the results of the paper.