Figure 1
From: Stochastic co-teaching for training neural networks with unknown levels of label noise
Noise transition matrices used in synthetic experiments. Transition matrices are equivalent to confusion matrices but they visualize noise distribution among classes. ias noise, or label-flipping noise, mimics observer-bias by substituting the true label with the label of the neighboring class. Uniform noise is achieved by replacing the true label with a randomly selected other label. Note that bias noise should always be <50% to ensure that the majority of instances remain correct.
