Learning with noisy labels aims to train neural networks with noisy labels.Current models handle instance-inde-pendent label noise(IIN)well;however,they fall short with real-world noise.In medical image classification...Learning with noisy labels aims to train neural networks with noisy labels.Current models handle instance-inde-pendent label noise(IIN)well;however,they fall short with real-world noise.In medical image classification,atypical samples frequently receive incorrect labels,rendering instance-dependent label noise(IDN)an accurate representa-tion of real-world scenarios.However,the current IDN approaches fail to consider the typicality of samples,which hampers their ability to address real-world label noise effectively.To alleviate the issues,we introduce typicality-and instance-dependent label noise(TIDN)to simulate real-world noise and establish a TIDN-combating framework to combat label noise.Specifically,we use the sample’s distance to decision boundaries in the feature space to repre-sent typicality.The TIDN is then generated according to typicality.We establish a TIDN-attention module to combat label noise and learn the transition matrix from latent ground truth to the observed noisy labels.A recursive algorithm that enables the network to make correct predictions with corrections from the learned transition matrix is proposed.Our experiments demonstrate that the TIDN simulates real-world noise more closely than the existing IIN and IDN.Furthermore,the TIDN-combating framework demonstrates superior classification performance when training with simulated TIDN and actual real-world noise.展开更多
Training with confident positive-labeled instances has received a lot of attention in Positive and Unlabeled(PU)learning tasks,and this is formally termed“Instance-Dependent PU learning”.In instance-dependent PU lea...Training with confident positive-labeled instances has received a lot of attention in Positive and Unlabeled(PU)learning tasks,and this is formally termed“Instance-Dependent PU learning”.In instance-dependent PU learning,whether a positive instance is labeled depends on its labeling confidence.In other words,it is assumed that not all positive instances have the same probability to be included by the positive set.Instead,the instances that are far from the potential decision boundary are with larger probability to be labeled than those that are close to the decision boundary.This setting has practical importance in many real-world applications such as medical diagnosis,outlier detection,object detection,etc.In this survey,we first present the preliminary knowledge of PU learning,and then review the representative instance-dependent PU learning settings and methods.After that,we thoroughly compare them with typical PU learning methods on various benchmark datasets and analyze their performances.Finally,we discuss the potential directions for future research.展开更多
基金funded by the National Natural Science Foundation of China,No.62371139the Science and Technology Commission of Shanghai Municipality,Nos.22ZR1404800 and 22DZ1100101.
文摘Learning with noisy labels aims to train neural networks with noisy labels.Current models handle instance-inde-pendent label noise(IIN)well;however,they fall short with real-world noise.In medical image classification,atypical samples frequently receive incorrect labels,rendering instance-dependent label noise(IDN)an accurate representa-tion of real-world scenarios.However,the current IDN approaches fail to consider the typicality of samples,which hampers their ability to address real-world label noise effectively.To alleviate the issues,we introduce typicality-and instance-dependent label noise(TIDN)to simulate real-world noise and establish a TIDN-combating framework to combat label noise.Specifically,we use the sample’s distance to decision boundaries in the feature space to repre-sent typicality.The TIDN is then generated according to typicality.We establish a TIDN-attention module to combat label noise and learn the transition matrix from latent ground truth to the observed noisy labels.A recursive algorithm that enables the network to make correct predictions with corrections from the learned transition matrix is proposed.Our experiments demonstrate that the TIDN simulates real-world noise more closely than the existing IIN and IDN.Furthermore,the TIDN-combating framework demonstrates superior classification performance when training with simulated TIDN and actual real-world noise.
基金supported by National Natural Science Foundation of China(62336003,12371510)NSF for Distinguished Young Scholar of Jiangsu Province(BK20220080).
文摘Training with confident positive-labeled instances has received a lot of attention in Positive and Unlabeled(PU)learning tasks,and this is formally termed“Instance-Dependent PU learning”.In instance-dependent PU learning,whether a positive instance is labeled depends on its labeling confidence.In other words,it is assumed that not all positive instances have the same probability to be included by the positive set.Instead,the instances that are far from the potential decision boundary are with larger probability to be labeled than those that are close to the decision boundary.This setting has practical importance in many real-world applications such as medical diagnosis,outlier detection,object detection,etc.In this survey,we first present the preliminary knowledge of PU learning,and then review the representative instance-dependent PU learning settings and methods.After that,we thoroughly compare them with typical PU learning methods on various benchmark datasets and analyze their performances.Finally,we discuss the potential directions for future research.
