We show that label noise exists in adversarial training. Such label noise is due to the mismatch between the true label distribution of adversarial examples and the label inherited from clean examples - the true label distribution is distorted by the adversarial perturbation, but is neglected by the common practice that inherits labels from clean examples. Recognizing label noise sheds insights on the prevalence of robust overfitting in adversarial training, and explains its intriguing dependence on perturbation radius and data quality.

We show that label noise exists in adversarial training. Such label noise is due to the mismatch between the true label distribution of adversarial examples and the label inherited from clean examples – the true label distribution is distorted by the adversarial perturbation, but is neglected by the common practice that inherits labels from clean examples. Recognizing label noise sheds insights on the prevalence of robust overfitting in adversarial training, and explains its intriguing dependence on perturbation radius and data quality. Guided by our analyses, we proposed a method to automatically calibrate the label to address the label noise and robust overfitting. Our method achieves consistent performance improvements across various models and datasets without introducing new hyper-parameters or additional tuning.

In this paper, the authors propose combining a knowledge distillation and GANs to improve the accuracy for multi-class classification. At the core, they demonstrate that combining these two approaches provides a better balance of sample efficiency and convergence to the ground truth distribution for improved accuracy. They claim two primary technical innovations (beyond combining these two approaches): using the Gumbel-Max trick for differentiability and having the classifier supervise the teacher (not just the teacher supervise the classifier). They argue that the improvements come from lower variance gradients and that the equilibrium of the minimax game is convergence to the true label distribution. The idea of combining these two perspectives is interesting, and both the theoretical arguments and the empirical results are compelling.

How to conduct teacher training for knowledge distillation is still an open problem. It has been widely observed that a best-performing teacher does not necessarily yield the best-performing student, suggesting a fundamental discrepancy between the current teacher training practice and the ideal teacher training strategy. To fill this gap, we explore the feasibility of training a teacher that is oriented toward student performance with empirical risk minimization (ERM). Our analyses are inspired by the recent findings that the effectiveness of knowledge distillation hinges on the teacher's capability to approximate the true label distribution of training inputs. We theoretically establish that the ERM minimizer can approximate the true label distribution of training data as long as the feature extractor of the learner network is Lipschitz continuous and is robust to feature transformations. In light of our theory, we propose a teacher training method SoTeacher which incorporates Lipschitz regularization and consistency regularization into ERM. Experiments on benchmark datasets using various knowledge distillation algorithms and teacher-student pairs confirm that SoTeacher can improve student accuracy consistently.

Label distribution learning can characterize the polysemy of an instance through label distributions. However, some noise and uncertainty may be introduced into the label space when processing label distribution data due to artificial or environmental factors. To alleviate this problem, we propose a \textbf{L}abel \textbf{C}orrelation \textbf{G}rid (LCG) to model the uncertainty of label relationships. Specifically, we compute a covariance matrix for the label space in the training set to represent the relationships between labels, then model the information distribution (Gaussian distribution function) for each element in the covariance matrix to obtain an LCG. Finally, our network learns the LCG to accurately estimate the label distribution for each instance. In addition, we propose a label distribution projection algorithm as a regularization term in the model training process. Extensive experiments verify the effectiveness of our method on several real benchmarks.

Here, we show that the robust overfitting shall be viewed as the early part of an epoch-wise double descent -- the robust test error will start to decrease again after training the model for a considerable number of epochs. Inspired by our observations, we further advance the analyses of double descent to understand robust overfitting better. In standard training, double descent has been shown to be a result of label flipping noise. However, this reasoning is not applicable in our setting, since adversarial perturbations are believed not to change the label. Going beyond label flipping noise, we propose to measure the mismatch between the assigned and (unknown) true label distributions, denoted as \emph{implicit label noise}. We show that the traditional labeling of adversarial examples inherited from their clean counterparts will lead to implicit label noise. Towards better labeling, we show that predicted distribution from a classifier, after scaling and interpolation, can provably reduce the implicit label noise under mild assumptions. In light of our analyses, we tailored the training objective accordingly to effectively mitigate the double descent and verified its effectiveness on three benchmark datasets.

The predictive performance of supervised learning algorithms depends on the quality of labels. In a typical label collection process, multiple annotators provide subjective noisy estimates of the "truth" under the influence of their varying skill-levels and biases. Blindly treating these noisy labels as the ground truth limits the accuracy of learning algorithms in the presence of strong disagreement. This problem is critical for applications in domains such as medical imaging where both the annotation cost and inter-observer variability are high. In this work, we present a method for simultaneously learning the individual annotator model and the underlying true label distribution, using only noisy observations. Each annotator is modeled by a confusion matrix that is jointly estimated along with the classifier predictions. We propose to add a regularization term to the loss function that encourages convergence to the true annotator confusion matrix. We provide a theoretical argument as to how the regularization is essential to our approach both for the case of single annotator and multiple annotators. Despite the simplicity of the idea, experiments on image classification tasks with both simulated and real labels show that our method either outperforms or performs on par with the state-of-the-art methods and is capable of estimating the skills of annotators even with a single label available per image.

In this paper, we consider a novel machine learning problem, that is, learning a classifier from noisy label distributions. In this problem, each instance with a feature vector belongs to at least one group. Then, instead of the true label of each instance, we observe the label distribution of the instances associated with a group, where the label distribution is distorted by an unknown noise. Our goals are to (1) estimate the true label of each instance, and (2) learn a classifier that predicts the true label of a new instance. We propose a probabilistic model that considers true label distributions of groups and parameters that represent the noise as hidden variables. The model can be learned based on a variational Bayesian method. In numerical experiments, we show that the proposed model outperforms existing methods in terms of the estimation of the true labels of instances.