How to model noise data? Hong Kong Baptist University's latest review paper on "label noise representation learning" comprehensively expounds the data, objective function and optimization strategy of

The article links ：https://arxiv.org/abs/2011.04406

Classical machine learning implicitly assumes that the labels of training data are sampled from a clean distribution , This may be too restrictive for real-world scenes . However , Statistical learning based methods may not be able to robustly train deep learning models under these noisy labels . therefore , Design labels - Noise means learning (Label-Noise Representation Learning, LNRL) Methods it is urgent to carry out robust training on the depth model with noise labels . In order to fully understand LNRL, We conducted a research . We first clarify from the perspective of machine learning LNRL Formal definition of . then , From the perspective of theoretical and empirical research , We find out why noisy labels affect the performance of depth models . On the basis of theoretical guidance , We will be different LNRL The method is divided into three directions . Under this unified classification , We have a comprehensive discussion on the advantages and disadvantages of different categories . what's more , We summarize robustness LNRL Basic components of , Can inspire new directions . Last , We proposed LNRL Possible research directions , Nu Skin dataset 、 Instance dependent LNRL And confrontation LNRL. We also look forward to LNRL Potential directions beyond , Such as characteristic noise learning 、 Prefer noise learning 、 Domain noise learning 、 Similar noise learning 、 Figure noise learning and demonstration noise learning .