A Derivation of D1 Denote the logit vector as x, we have p j = e

Without zero-mean constraint, the training becomes unstable. For GLC, we first train 40 epochs to estimate the label corruption matrix and then train another 40 epochs to evaluate its performance. Since Co-teach uses two models, each model is trained for 40 epochs for a fair comparison. We use one V100 GPU for all the experiments. Table 6: Ratio of increased class-level weights under the imbalance setting.weight/class