Iterative Learning with Open-set Noisy Labels

Abstract

Large-scale datasets possessing clean label annotations are crucial for training Convolutional Neural Networks (CNNs). However, labeling large-scale data can be very costly and error-prone, and even high-quality datasets are likely to contain noisy (incorrect) labels. Existing works usually employ a closed-set assumption, whereby the samples associated with noisy labels possess a true class contained within the set of known classes in the training data. However, such an assumption is too restrictive for many applications, since samples associated with noisy labels might in fact possess a true class that is not present in the training data. We refer to this more complex scenario as the open-set noisy label problem and show that it is nontrivial in order to make accurate predictions. To address this problem, we propose a novel iterative learning framework for training CNNs on datasets with open-set noisy labels. Our approach detects noisy labels and learns deep discriminative features in an iterative fashion. To benefit from the noisy label detection, we design a Siamese network to encourage clean labels and noisy labels to be dissimilar. A reweighting module is also applied to simultaneously emphasize the learning from clean labels and reduce the effect caused by noisy labels. Experiments on CIFAR-10, ImageNet and real-world noisy (web-search) datasets demonstrate that our proposed model can robustly train CNNs in the presence of a high proportion of open-set as well as closed-set noisy labels.

NLNL: Negative Learning for Noisy Labels

“…Other approaches include jointly modeling labels and worker quality [15], creating a robust method to learn in open-set noisy label situations [34] and attempting to prune the correct samples [6,2,24]. Ding et al [2] suggested pruning the correct samples based on softmax outputs.…”

“…Regarding noisy data, while PL provides CNN the wrong information (red balloon), with a higher chance, NL can provide CNN the correct information (blue balloon) because a dog is clearly not a bird. approaches address this problem by applying a number of techniques and regularization terms along with Positive Learning (PL), a typical supervised learning method for training CNNs that "input image belongs to this label" [6,2,34,20,3,39,26,30,22,33,21]. However, when the CNN is trained with images and mismatched labels, wrong information is being provided to the CNN.…”

Deep Multi-Modal Object Detection and Semantic Segmentation for Autonomous Driving: Datasets, Methods, and Challenges

“…The network is much more robust against labeling errors when they are randomly distributed, compared to biased labeling from the use of a deterministic pre-labeling. Training networks with erroneous labels is further studied in [188]- [191]. The impact on weak or erroneous labels on the performance of deep learning based semantic segmentation is investigated in [192], [193].…”

Deep Graph Regularized Learning for Binary Classification

“…Compared to the previous graph-based classifiers [18], [20], [21], [23], [29], [40], [41], [45], [46], by adopting edge convolution, iteratively updating graph and operating GLR, we learn a deeper feature representation, and assign the degree of freedom for learning the underlying data structure. Given noisy training labels, in contrast to the classical robust DNN-based classifiers [4], [7], [15], [16], [35], [39], we bring together the regularization benefits of GLR and the benefits of the proposed loss functions to perform more robust deep metric learning. We further adopt a rank-sampling strategy to find those training samples with high predictive performance that benefits inference.…”

“…Θ is an edge attention activation function (see (7) for the particular function we used) that estimates how much attention should be given to each edge and Y r = {Ẏ r = [−1, 1] M , [−1, 1] N −M } is the restored classifier signal obtained via (3) starting from the classifier signal in the previous iteration, Y r−1 . π ψi,ψj is the amount of attention, i.e., edge loss weights, assigned to the edge connecting vertices ψ i and ψ j .…”