Learning to rectify for robust learning with noisy labels

Label noise significantly degrades the generalization ability of deep models in applications. Effective strategies and approaches (e.g., re-weighting or loss correction) are designed to alleviate negative impact label when training a neural network. Those existing works usually rely on pre-specified architecture manually tuning additional hyper-parameters. In this paper, we propose warped probabilistic inference (WarPI) achieve adaptively rectifying procedure for classification network within meta-learning scenario. contrast deterministic models, WarPI is formulated as hierarchical model by learning an amortization meta-network, which can resolve sample ambiguity be therefore more robust serious noise. Unlike approximated weighting function directly generating weight values from losses, our meta-network learned estimate vector input logits labels, has capability leveraging sufficient information lying them. The provides effective way rectify network, demonstrating significant improvement ability. Besides, modeling latent variable seamlessly integrated into SGD optimization We evaluate four benchmarks with noisy labels new state-of-the-art under variant types. Extensive study analysis also demonstrate effectiveness model.