Learning to rectify for robust learning with noisy labels

Highlights：

• Our WarPI is the first probabilistic model to resolve label noise within the meta-learning scenario.

• We design a powerful amortized meta-network to estimate the distribution of the rectifying vector from the input of labels and predicted vector.

• WarPI can be directly integrated into the training of the prediction network, demonstrating favorable effectiveness to learn from noisy labels.

• WarPI achieves the new state-of-the-art on four challenging benchmarks under variant noise types.