Open-set Label Noise Can Improve Robustness Against Inherent Label Noise
Extracted source text · arXiv 2106.10891
\documentclass{article}
\PassOptionsToPackage{numbers, compress}{natbib}
\usepackage[final]{neurips_2021}
% to compile a preprint version, e.g., for submission to arXiv, add add the
% [preprint] option:
% \usepackage[preprint]{neurips_2021}
% to compile a camera-ready version, add the [final] option, e.g.:
% \usepackage[final]{neurips_2021}
% to avoid loading the natbib package, add option nonatbib:
% \usepackage[nonatbib]{neurips_2021}
\usepackage[utf8]{inputenc} % allow utf-8 input
\usepackage[T1]{fontenc} % use 8-bit T1 fonts
% \usepackage{hyperref} % hyperlinks
\usepackage{url} % simple URL typesetting
\usepackage{booktabs} % professional-quality tables
\usepackage{amsfonts} % blackboard math symbols
\usepackage{nicefrac} % compact symbols for 1/2, etc.
\usepackage{microtype} % microtypography
\usepackage{xcolor} % colors
\usepackage{graphicx}
\usepackage{amsmath}
\usepackage{amssymb}
% \usepackage{caption}
\usepackage{subcaption}
\usepackage{algorithm}
\usepackage{algorithmic}
\usepackage{multirow}
\usepackage{booktabs}
% \usepackage{wrapfig}
\usepackage{wrapfig}
% \usepackage{hyperref}
\usepackage[toc,page]{appendix}
\usepackage{amsthm}
\renewcommand{\algorithmicrequire}{ \textbf{Input:}} %Use Input in the format of Algorithm
\newtheorem{prop}{Proposition}
%%% Specially configured color
\usepackage{xcolor}
\definecolor{mydarkred}{rgb}{0.6,0,0}
\definecolor{mydarkgreen}{rgb}{0,0.6,0}
\usepackage[colorlinks,
linkcolor=mydarkred,
citecolor=mydarkgreen]{hyperref}
\title{Open-set Label Noise Can Improve Robustness Against Inherent Label Noise }
% \title{Open-set noise can improve the robustness against label noise}
%
%How does Out of Distribution Samples Help Generalization against Label Noise
%Understanding the effect of Out of Distribution Samples on Generalization and robustness
%Does Open-set Noise really harm
%
% The \author macro works with any number of authors. There are two commands
% used to separate the names and addresses of multiple authors: \And and \AND.
%
% Using \And between authors leaves it to LaTeX to determine where to break the
% lines. Using \AND forces a line break at that point. So, if LaTeX puts 3 of 4
% authors names on the first line, and the last on the second line, try using
% \AND instead of \And before the third author name.
% \author{%
% David S.~Hippocampus\thanks{Use footnote for providing further information
% about author (webpage, alternative address)---\emph{not} for acknowledging
% funding agencies.} \\
% Department of Computer Science\\
% Cranberry-Lemon University\\
% Pittsburgh, PA 15213 \\
% \texttt{hippo@cs.cranberry-lemon.edu} \\
% }
\author{Hongxin Wei$^{1}$ \quad Lue Tao$^{2,3}$ \quad Renchunzi Xie$^{1}$\thanks{Corresponding author.} \quad Bo An$^{1}$\\
$^1$School of Computer Science and Engineering, Nanyang Technological University, Singapore \quad\\
$^2$ College of Computer Science and Technology, \\
Nanjing University of Aeronautics and Astronautics, China\\
$^3$ MIIT key Laboratory of Pattern Analysis and Machine Intelligence, China\\
{\tt\small \{hongxin001,xier0002\}@e.ntu.edu.sg \quad tlmichael@nuaa.edu.cn \quad boan@ntu.edu.sg}
}
\begin{document}
\maketitle
\begin{abstract}
% Deep neural networks always suffer from noisy labels
% Deep neural network could easily memorize noisy labels due to its sufficient capacity.
% To mitigate this issue, a popular idea is designing a robust loss function or regularization to alleviate the overfitting to noisy labels. However, existing algorithms only consider the original training dataset, ignoring the possibility of utilizing auxiliary dataset for further improvement.
% Although many robust algorithms have been proposed to mitigate this issue, they only consider the original training dataset, ignoring the possibility of utilizing auxiliary dataset for further improvement.
% In this work, we are interested in improving robustness to noisy labels with an open-set auxiliary dataset.
% This phenomenon could be intuitively explained that fitting open-set label noises consumes extra capacity of the network, thereby reducing the memorization of inherent noisy labels.
% Through the lens of SGD noise, we show that the noises induced by our method are random-direction, conflict-free and biased, which may help the model converge to a flat minimum and maintain good stability. Besides, the biases of the noises enforce the model to produce conservative predictions on Out-of-Distribution instances.
% Furthermore, we analyze the effect of ODNL from the lens of SGD noise and formalize the differences and connections between ODNL and related methods.
% Extensive experimental results on corrupted data from benchmark datasets demonstrate that the proposed method could consistently enhance the performance of many state-of-the-art approaches. In addition, we observe ODNL could also achieve significant improvement
% significantly improve Out-of-Distribution detection performance
Learning with noisy labels is a practically challenging problem in weakly supervised learning. In the existing literature, open-set noises are always considered to be poisonous for generalization, similar to closed-set noises. In this paper, we empirically show that open-set noisy labels can be non-toxic and even benefit the robustness against inherent noisy labels. Inspired by the observations, we propose a simple yet effective regularization by introducing Open-set samples with Dynamic Noisy Labels (ODNL) into training. With ODNL, the extra capacity of the neural network can be largely consumed in a way that does not interfere with learning patterns from clean data. Through the lens of SGD noise, we show that the noises induced by our method are random-direction, conflict-free and biased, which may help the model converge to a flat minimum with superior stability and enforce the model to produce conservative predictions on Out-of-Distribution instances. Extensive experimental results on benchmark datasets with various types of noisy labels demonstrate that the proposed method not only enhances the performance of many existing robust algorithms but also achieves significant improvement on Out-of-Distribution detection tasks even in the label noise setting.
% Due to the sufficient capacity, deep neural networks could easily memorize closed-set noisy labels, as well as open-set noisy labels.
% Deep neural networks could easily memorize noisy labels because of its sufficient capacity. Closed-set and open-set noisy labels are extensively explored by existing work,
% It has been shown that open-set noisy labels are less harmful than closed-set noisy labels, but little is known about the interaction between the two types of noisy labels during training. In this work, we empirically show that controlled open-set noises can be ``non-toxic poison'' and even benefit the robustness against inherent noisy labels. This phenomenon could be intuitively explained that fitting open-set noise consumes extra capacity of the network, thereby reducing the memorizing inherent noisy labels. Inspired by the observations, We propose a simple but effective regularization by introducing open-set samples with dynamic noise labels. In this way, extra capacity of the network is fully consumed in a way that does not interfere with learning patterns from clean data.
%Closed-set noisy labels has been shown to be harmful for generalization and existing work focus on reducing the their impact in training, by sample weighting, selection or noise transition matrix. Open-set noisy labels are treated in a similar way.
% Noisy labels could be easily memorized due to sufficient capacity of deep neural networks
% Open-set samples has been widely studied in the context of open-set noisy la+
% bel, open-set semi supervised learning and out-of-distribution detection. Despite its popularity in the literature, where existing work mainly focuses on filtering out open-set samples from train sets or test sets, little is known about their influence on generalization and robustness. In this paper, we show that randomly labelled open-set noises slightly harm the generalization of deep learning, but improve the robustness against label noise. As the number of open-set noise increases, the negative effect on generalization is reduced.
% More surprisingly, the robustness against noisy labels are improved.
% Inspired by the observations, we propose a new method to improve robustness on noisy labels with limited number of open-set samples, through a simple modification. To be specific, we randomly select a number of open-set samples in each iteration and assign labels for these samples with uniform distribution. Then we calculate the cross entropy loss given the open-set samples and random labels. In this way,.
% Extensive experimental results on corrupted data from benchmark datasets including MNIST, CIFAR-10, CIFAR-100 and Clothing1M demonstrate that our regularization could consistently improve the performance of many state-of-the-art approaches in robustness against label noise. Moreover, our regularization could also improve performance on OOD detection task, especially under the setting of noisy labels.
\end{abstract}
\section{Introduction}
% Deep Neural Networks (DNNs) achieve remarkable success on various tasks, and most of them are trained in a supervised manner, which heavily relies on a large number of training instances with accurate labels.
The success of deep neural networks (DNNs) heavily relies on a large number of training instances with fully accurate labels \cite{he2016deep}.
In real-world applications, it is expensive and time-consuming to collect such large-scale datasets.
% with high-quality annotations.
To alleviate this problem, some surrogate methods are commonly used to improve labelling efficiency, such as online queries \cite{blum2003noise} and crowdsourcing \cite{yan2014learning}. These cheap but imperfect methods usually suffer from unavoidable noisy labels that can be easily memorized by Deep Neural Networks (DNNs), leading to poor generalization performance \cite{pmlr-v70-arpit17a, zhang2016understanding}. Therefore, designing robust algorithms against noisy labels is of great practical importance.
%Although these cheap but imperfect methods provide the possibility to scale the acquisition of training labels,
% Although regularization methods have been widely studied to improve the generalization performance in deep learning, like data augumentation \cite{shorten2019survey}, weight decay \cite{krogh1992simple} and dropout \cite{srivastava2014dropout}, these traditional methods are still prone to the overfitting issue of noisy labels. To alleviate this problem, increasing interests have been drawn in exploiting robust loss functions and regularization methods for training deep models against label noise. For example, PHuber
% Label noise can be separated into two categories: closed-set and open-set \cite{wang2018iterative}. The closed-set noise occurs when instances with noisy labels have true class labels within the noisy label set. Oppositely, the open-set noise occurs when instances with noisy labels have some true class labels outside the noisy label set.
Label noise can be separated into two categories: 1) closed-set noise, where instances with noisy labels have true class labels within the noisy label set \cite{han2018co, jiang2017mentornet, patrini2017making, reed2014training, van2015learning, wei2020combating}. 2) open-set noise, where instances with noisy labels have some true class labels outside the noisy label set \cite{lee2019robust, sachdeva2021evidentialmix, wang2018iterative, xia2020extended}.
In the existing literature, open-set noises are always considered to be harmful to the training of DNNs like closed-set noises \cite{lee2019robust, wang2018iterative, sachdeva2021evidentialmix, xia2020extended}. For example, ILON \cite{wang2018iterative} reduces the effect of open-set noises by using noisy label detector and contrastive loss to pull away noisy samples from clean samples. Extended T \cite{xia2020extended} uses an extended transition matrix to mitigate the impacts of open-set noises and EvidentialMix \cite{sachdeva2021evidentialmix} filters out the open-set examples by modelling the loss distribution. Therefore, there arises a question to be answered: \emph{Do open-set noises always play a negative role in the training?}
% Overall, all these methods are designed to deal with
% and EvidentialMix \cite{sachdeva2021evidentialmix} regard the mixed closed-set and open-set label noise as a more realistic and challenging problem.
% The former and the latter
% In this work, we are interested in improving robustness against noisy labels with an unlabelled open-set auxiliary dataset that is easy to get in real-world applications \cite{hendrycks2019oe}. To avoid confusion with noises from the auxiliary dataset, we use \textit{inherent noisy labels} to denote those from the original training dataset.
% Although open-set auxiliary dataset is previously studied in Out-of-Distribution (OOD) Detection tasks \cite{hendrycks2019oe},
% The closed-set label noise occurs when instances with noisy labels have true class labels within the noisy label set.
% The effect of closed-set label noise has been extensively explored by existing work \cite{chen2019understanding,zhang2016understanding}.
% The sufficient capacity of DNNs make it easy to perfectly fit the training set in any case, even with totally random labels. In particular, DNNs would memorize easy instances first, and gradually adapt to hard examples as the number of epochs increases. Therefore, closed-set noisy labels would deteriorate the learning performance of DNNs because the learned model will eventually memorize these wrongly given labels.
% Oppositely, the open-set label noise occurs when instances with noisy labels have some true class labels outside the noisy label set, which are also known as Out-of-Distribution (OOD) samples.
% Open-set label noise are also studied in some work. In addition, there also exist several related research directions studying the OOD samples, such as open-set semi-supervised learning and OOD detecion.
% However, all these methods are designed to reduce the impact of open-set label noise or directly filter out OOD samples, assuming they are "poison" for the training.
% Little is known about how open-set label noise or OOD samples affect the generalization and robustness of neural network.
% intentionally generated
% To avoid confusion, we use \textit{inherent noisy labels} to denote noisy labels from the original training dataset.
In this work, we answer this question by experiments with an open-set auxiliary dataset. We empirically show that additional open-set noises can be harmless to generalization if the number of open-set noises is sufficiently large. More surprisingly, we observe that the open-set noises can even benefit the robustness of neural networks against noisy labels from the original training dataset, while the closed-set counterpart could not. Based on the ``insufficient capacity'' hypothesis \cite{pmlr-v70-arpit17a}, an intuitive explanation for this phenomenon is that increasing the number of open-set auxiliary samples slows down the fitting of inherent noises. In other words, fitting open-set label noises consumes extra capacity of the network, thereby reducing the memorization of inherent noisy labels. The details of experimental results are presented in Subsection \ref{motivation}.
% To avoid confusion with noises from the auxiliary dataset, we use \textit{inherent noisy labels} to denote those from the original training dataset.
% It has been shown that network would first learn patterns from clean data, then turns to memorize data with noisy labels. This phenomenon could be intuitively explained by insufficient capacity: Increasing the number of examples while keeping representation capacity fixed would increase the time needed to memorize the training set.
% The former phenomenon is intuitive but the latter one is counter-intuitive.
% As the sample size of open-set label noise increases, fitting each OOD samples would be a more difficult task in the training. Therefore, it is intuitive that the toxicity of open-set label noise would be reduced on the generalization. On the other hand, the ``non-toxic poison'' brings hard examples with high losses to the training process, thereby reducing the overfitting to the noisy labels in training dataset. In this way, the ``non-toxic poison'' becomes an ``antidote'' to another ``poison''. Therefore, there arises a question to be answered: Can we design an ``impossible task" to reduce overfitting to label noise?
%to consistently improve existing methods for robustness against label noise.
% Inspired by the above observations, we propose a simple yet effective regularization by introducing Open-set samples with Dynamic Noisy Labels (ODNL) into training.
Inspired by the above observations, we propose an effective and complementary method to improve robustness against noisy labels. The high-level idea is to introduce Open-set samples with Dynamic Noisy Labels (ODNL) into training.
In each epoch, we assign random labels to instances of open-set auxiliary dataset by uniformly sampling from the label set. In other words, the label of each open-set sample is independently random and is constantly changing over training epochs. In this manner, ODNL consistently produce ``benign hard examples'' to consume the extra representation capacity of neural networks, thereby preventing the neural network from overfitting inherent noisy labels.
To further understand the effect of ODNL, we provide an analysis from the lens of SGD noise \cite{chen2021noise, wu2020noisy}, and formalize the differences and connections between our regularization and related methods. From the analysis, we show that the SGD noises induced by our method are 1) random-direction, helping the model converge to a flat minimum; 2) conflict-free, leading to good stability for the training; 3) biased,
enforcing the model to produce more conservative predictions on OOD examples. In this manner, our method not only improves the robustness against noisy labels, but also provides benefits for OOD detection performance, which is also essential for the deployment of deep learning in safety-critical applications.
% random, conflict-free, biased: flat minima, stable, improve OOD
% 1) dynamic noisy labels help escape from local minima by providing noisy gradients with ``random direction''. 2) open-set samples are utilized to generate the noisy gradients so that the noises is independently random to the original gradients w.r.t parameters of network. Otherwise, the noises might be conflicted with the original gradients, leading to unstable training. 3) the biases of the noises can improve predictive uncertainty and model calibration, especially on OOD samples.
% In each batch, we randomly select a number of OOD samples and assign them random labels with discrete uniform distribution.
% In particular, the random label for each OOD sample is changed over epochs.
% In standard training, these randomly labelled OOD samples could be directly incorporated into the training batch for training the classifier. For existing robust label-noise learning methods, we can add them into the training of DNNs in the form of regularization with a standard loss function, such as cross-entropy loss for multi-class classification. In this way, we create an ``impossible mission'' where non-toxic samples could never be fitted, thereby reducing overfitting in noisy labels.
To the best of our knowledge, we are the first to explore the benefits of open-set auxiliary dataset in learning with noisy labels. To verify the efficacy of our regularization, we conduct extensive experiments on both simulated and real-world noisy datasets, including CIFAR-10, CIFAR-100 and Clothing1M datasets. In summary, the proposed method can be easily adapted for a wide range of existing algorithms to prevent overfitting noisy labels.
Furthermore, experimental results validate that our regularization could also achieve impressive performance for detecting OOD examples, even if the labels of training dataset are noisy.
% while OE could not provide improvement in robustness against label noise. Experimental results show that our proposed method could achieve better performance on both prediction and OOD detection tasks
% Compared to open-set samples, ones with farther distance to the original sample distribution would cause less damage to generalization and make more improvements in robustness against close-set label noise.
\section{Problem setting and motivation}
% In this section, we introduce a simple yet effective method for better robustness against label noise. First, we briefly introduce the problem setting of learning from training data with noisy labels. Then we start in \ref{} with an observational study highlighting the effect of open-set noise examples with different sample size for generalization and robustness against label noise. Motivated by the observations, we propose to use open-set samples with dynamic noisy labels to alleviate overfitting to the corrupted labels in training data. Finally, in section \ref{}, we propose a simple yet effective regularization to consistently improve existing algorithms for robustness against label noise.
In this section, we first briefly introduce the problem setting of learning from training data with noisy labels. Then we start in Subsection \ref{motivation} with an observational study highlighting the effect of open-set noisy examples from the auxiliary dataset on generalization and robustness against noisy labels.
\subsection{Preliminaries: learning with inherent noisy labels} \label{preliminaries}
In this work, we consider multi-class classification with $k$ classes. Let $\mathcal{X}\subset\mathbb{R}^{d}$ be the feature space and $\mathcal{Y} = \{1, \ldots, k\}$ be the label space, we suppose the training dataset with $N$ samples is given as $\{\boldsymbol{x}_i, y_i\}^N_{i=1}$, where $\boldsymbol{x}_i \in \mathcal{X}$ is the $i$-th instance sampled from a certain data-generating distribution $\mathcal{D}_{\mathrm{train}}$ in an i.i.d. manner and $y_i \in \{1, \ldots, k\}$ represents its observed label that might be different from the ground truth. To prevent confusion, the noisy labels in the training dataset are referred to as \textit{inherent noisy labels}. A classifier is a function that maps the feature space to the label space $f: \mathcal{X} \rightarrow \mathbb{R}^{k}$ with trainable parameter $\boldsymbol{\theta} \in \mathbb{R}^{p}$. In this paper, we consider the common case where the function $f$ is a DNN with the softmax output layer. The objective function of standard training can be represented as a cross-entropy loss:
\begin{equation}
\mathcal{L}_{CCE}=\mathbb{E}_{\mathcal{D}_{\mathrm{train}}}\left[\ell\left(f(\boldsymbol{x} ; \boldsymbol{\theta}), y\right)\right] \simeq -\frac{1}{N} \sum_{i=1}^{N} \boldsymbol{e}^{y_i} \log f\left(\boldsymbol{x}_{i} ; \boldsymbol{\theta}\right),
\end{equation}
\noindent where $f\left(\boldsymbol{x}_{i} ; \boldsymbol{\theta}\right)$ denotes the output of the classifier $f$ with $\boldsymbol{\theta}$ for $\boldsymbol{x}_i$ and $\boldsymbol{e}^{y_i} := (0, \ldots, 1, \ldots, 0)^{\top} \in\{0,1\}^{k}$ is a one-hot vector and only the $y_i$-th entry of $\boldsymbol{e}^{y_i}$ is 1.
In addition to the training data, we consider there is an unlabelled auxiliary dataset consisting of open-set instances $\{\widetilde{\boldsymbol{x}}_i\}^M_{i=1}$. Specifically, open-set instances are sampled from $\mathcal{D}_{\mathrm{out}}$, disjoint
from $\mathcal{D}_{\mathrm{train}}$. Therefore, they are also called Out-of-Distribution (OOD) instances. In real applications, it is easy to get such an auxiliary dataset, which is commonly used in OOD detection tasks \cite{hendrycks2019oe, papadopoulos2021outlier}.
% To involve the open-set noisy dataset into training, a simple method is to directly train on instances sampled from the union of original dataset and auxiliary dataset. However, the
% increased open-set noisy examples would further dilute the gradients of training examples from original dataset in each iteration. To avoid unnecessary impacts, we sample a fixed number of open-set instances over experiments with various sample sizes in each iteration. Then the objective function becomes:
%\mathbb{E}_{\left(\mathcal{D}_{train} \ \cup \ \mathcal{D}_{out}\right)}\left[\mathcal{L}\left(f(\boldsymbol{x} ; \boldsymbol{\theta}), y_{\boldsymbol{x}}\right)\right]
% \begin{equation}
% \mathcal{L}_{CCE} = -\frac{1}{n+m} \left(\sum_{i=1}^{n} \boldsymbol{e}^{y_i} \log f\left(\boldsymbol{x}_{i} ; \boldsymbol{\theta}\right) + \sum_{i=1}^{m} \boldsymbol{e}^{\widetilde{y}_i} \log f\left(\widetilde{\boldsymbol{x}}_{i} ; \boldsymbol{\theta}\right)\right)
% \end{equation}
% \noindent where $n$ and $m$ denote the batch sizes of original dataset and open-set noisy dataset respectively.
\begin{figure}
\centering
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{img/size_fig_clean.pdf}
\caption{Learning with Clean data}
\label{fig:size_clean}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{img/size_fig_noise.pdf}
\caption{Learning with inherent noises}
\label{fig:size_noise}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{img/cifar10_params_loss.pdf}
\caption{Losses of inherent noises}
\label{fig:in_noise_loss}
\end{subfigure}
\caption{Average test accuracy (\%) on CIFAR-10 with (a) clean labels (b) inherent noisy labels under different auxiliary dataset sizes. We use ``open-set'' and ``closed-set'' to denote two types of auxiliary datasets. The results of standard training are also reported for comparison. All experiments are repeated five times and the standard deviations are presented as error bars. (c) Training losses of inherent noises around the 140th epoch under different auxiliary dataset sizes (K). }
\label{fig:motivation}
% \vspace{-10pt}
\end{figure}
\subsection{Motivation: the role of open-set label noise in generalization and robustness} \label{motivation}
% We first motivate our method by performing an observation study focusing on how the sample size $M$ of the given open-set auxiliary dataset affects the generalization performance of DNNs under different settings. To demonstrate the importance of the sample size, we run standard training with different-sized auxiliary dataset. As the sample size of the auxiliary dataset increases, memorizing noisy samples in the auxiliary dataset would require more representational capacity of network \cite{pmlr-v70-arpit17a}.
% \noindent\textbf{Experimental setup.}
To analyze the effects of open-set noisy labels, we build an open-set noisy dataset with the unlabelled auxiliary dataset. For each open-set instance, we assign a fixed random label $\widetilde{y}_i$ uniformly sampled from the label set $\{1, \ldots, k\}$. In each iteration, we concatenate the two batches that are sampled from the training dataset and the open-set noisy dataset respectively, then perform standard training.
% To avoid confusion with noises from the auxiliary dataset, we use \textit{inherent noisy labels} to denote those from the original training dataset.
In addition, we conduct controlled experiments by replacing the open-set dataset with a closed-set dataset to verify the effectiveness of open-set noisy dataset during the training. To demonstrate the importance of the sample size of the given auxiliary dataset, we compare the performance of training with different-sized auxiliary datasets, built by randomly sampling from the original auxiliary dataset.
%As the sample size of the auxiliary dataset increases, memorizing noisy samples in the auxiliary dataset would require more representational capacity of network.
Throughout this subsection, we perform training with WRN-40-2 \cite{zagoruyko2016wide} on CIFAR-10 \cite{krizhevsky2009learning} using standard cross-entropy loss. For experiments under inherent label noise, we use symmetric noise with 40\% noisy rate. For auxiliary dataset, we use 300K Random Images\footnote{The dataset is published on \url{https://github.com/hendrycks/outlier-exposure}.} \cite{hendrycks2019oe} as the open-set auxiliary dataset and CIFAR-5m\footnote{The dataset is published on \url{https://github.com/preetum/cifar5m}.} \cite{ho2020denoising, nakkiran2020deep} as the closed-set auxiliary dataset. Specifically,
% we remove all examples of 80 Million Tiny Images that appear in the CIFAR-10/100 datasets,
all images that belong to CIFAR classes are removed in 300K Random Images so that $\mathcal{D}_{\mathrm{train}}$ and $\mathcal{D}_{\mathrm{out}}$ are disjoint. More training details can be found in Appendix \ref{app:exp_settings}.
% with 6 million synthetic CIFAR-10-like images.
%We perform training with WideResNet on CIFAR-10 using standard cross-entropy loss. The network is trained for 200 epochs using SGD with a momentum of 0.9, a weight decay of 0.0005. In each iteration, both the batch sizes of original dataset and open-set noisy dataset are set as 128. For experiments under inherent label noise, we use symmetric noise with 40\% noisy rate. We set the initial learning rate as 0.1, and reduce it by a factor of 10 after 80 and 140 epochs.
% \noindent\textbf{Effect of different labels.} To show the role of labels for open-set noise in the generalization performance, we run standard training on clean training data and OOD samples with different labelling method. Figure summarizes the result of the experiment. We observe when open-set samples are with random labels sampled from discrete uniform distribution, it cause the least negative impact on generalization.
\noindent\textbf{Learning with clean labels.} Fig.~\ref{fig:size_clean} presents the results on clean training set of CIFAR-10, training with the open/closed-set noisy auxiliary dataset. The results show that open-set noisy labels from the auxiliary dataset become harmless gradually with the increase of the sample size of the auxiliary dataset, while closed-set noisy labels consistently deteriorate the generalization performance. Furthermore, we observe that closed-set noisy labels result in unstable training shown by the large standard deviations in the figure. It is worthy to note that open-set noisy labels do not encounter this issue.
%the test accuracy of training with open-set noisy labels are consistently higher than those of training with closed-set noisy labels, which proves that open-set label noise causes less damage to generalization than closed-set label noise. With the increase of the sample size of auxiliary dataset, open-set label noise tends to be harmless, while closed-set label noise keeps playing a negative role on generalization performance. In summary, these results show that open-set label noise could be "non-toxic poison" compared to closed-set label noise.
\noindent\textbf{Learning with inherent noisy labels.} Fig.~\ref{fig:size_noise} presents the results on CIFAR-10 with inherent noisy labels, training with the open/closed-set noisy auxiliary dataset. Surprisingly, we can observe that as we increase the sample size of the open-set noisy dataset, the open-set noises can even improve the generalization performance beyond standard training without the noisy auxiliary dataset. In this case, closed-set noisy auxiliary dataset still plays a negative role in the training.
\noindent \textbf{An intuitive interpretation}. From the results, one can hypothesize that open-set noisy labels in the auxiliary dataset would be harmless and can even benefit the robustness against inherent noisy labels, if the sample size of the auxiliary dataset is large enough. This phenomenon is somewhat counter-intuitive since one would expect that open-set noisy labels should have been ``poison'' to the training of network, like closed-set noises \cite{lee2019robust, sachdeva2021evidentialmix, wang2018iterative, xia2020extended}. Here we provide an intuitive interpretation from ``insufficient capacity'' \cite{pmlr-v70-arpit17a}: increasing the number of examples while keeping representation capacity fixed would increase the time needed to memorize the data set. Hence, the larger the size of auxiliary dataset is, the more time it needs to memorize the open-set noises in the auxiliary dataset as well as the inherent noises in the training set, relative to clean data. This interpretation is supported by our plot of training losses of inherent noises with different-sized auxiliary datasets in Fig.~\ref{fig:in_noise_loss}, where increasing the number of open-set auxiliary samples slows down the fitting of inherent noises.
%memorizing the open-set noises in axiliary dataset would consume more capacity of network.
%increase the time to memorize noisy labels by consume extra capacity of network
% It has been shown that network would first learn patterns from clean data, then turns to memorize data with noisy labels. This phenomenon could be intuitively explained by insufficient capacity: Increasing the number of examples while keeping representation capacity fixed would increase the time needed to memorize the training set. For example, given a network that could just memorize $n$ samples with inherent noisy labels after learning clean patterns, now we add $m$ open-set samples with random labels into training, the network would randomly memorize $n$ samples from the $m+n$ samples with noisy labels. As increasing the size $m$ of open-set samples, the number of the memorized data with inherent noisy labels would decrease. From the former experiment, we know that open-set samples with random labels are non-toxic as the sample size increases. Therefore, we can intuitively understand why open-set samples with random labels could improve robustness against inherent noisy labels.
% It has been shown that DNNs learn simple patterns first, before memorizing \cite{pmlr-v70-arpit17a}.
% From the observations, Learning with the auxiliary dataset should not interfere with learning the real data.
\section{Method: Open-set regularization with Dynamic Noisy Labels}
%high level idea
Motivated by the previous observations, we propose to utilize open-set auxiliary dataset to further improve the robustness against inherent noisy labels. This approach can be considered as an auxiliary task that continuously consumes the extra capacity of neural networks but does not interfere with learning patterns from clean data. Such auxiliary tasks will prevent deep networks from overfitting noisy data.
% we propose a simple yet effective method that utilizes open-set noisy dataset for further improving the robustness against label noise. The high level idea is to make DNN stuck in memorizing ``non-toxic" data after learning the patterns from clean data, hence the network becomes robust to noisy labels. Recall that DNNs learn simple patterns first, before memorizing noisy labels that leads to the poor generalization performance. From the experimental results in \ref{motivation}, we can conjecture that memorizing open-set noise is more difficult than learning patterns from clean labels, and at least not more difficult than memorizing closed-set noisy labels.
%turn to the motivation of Dynamic Noisy label
\noindent \textbf{Dynamic Noisy Labels.} To maintain the effectiveness of the auxiliary task during training, an ideal method is to use an auxiliary dataset with unlimited instances based on ``insufficient capacity'' \cite{pmlr-v70-arpit17a}. In this way, the network would always try to fit a stream of new instances sequentially in an online learning setting.
% by optimize predictive models over a stream of data instances sequentially
%The logic behind is increasing the number of examples while keeping representational capacity fixed would increases the time needed to memorize the training set \cite{pmlr-v70-arpit17a}.
However, it is unrealistic to provide such an auxiliary dataset in real scenarios.
%What is Dynamic Noisy labels
To relieve the requirement on sample size, we propose an alternative approach called Dynamic Noisy Labels. For each open-set instance $\widetilde{\boldsymbol{x}}_i$, a noisy label $\widetilde{y}_i$ is uniformly sampled from the label set $\{1,\ldots,k\}$ independently, regardless of the instance $\widetilde{\boldsymbol{x}_i}$.
% drawn from a discrete uniform distribution $\mathcal{U}_{k}$, regardless of the instance $\widetilde{\boldsymbol{x}_i}$. In other word, the noisy labels are uniformly sampled from the label set independently.
More importantly, the generated labels are not fixed and would be changed dynamically over training epochs. In this way, the auxiliary task becomes an ``impossible mission", continuously consuming the extra capacity of network. Without Dynamic Noisy Labels, the effectiveness of the auxiliary task would be degraded gradually during training, because deep network can easily memorize the finite open-set samples, even though the labels of those samples are randomly generated \cite{pmlr-v70-arpit17a,zhang2016understanding}.
%The effect of Dynamic Noisy labels
\noindent\textbf{Training Objective.} For the original training dataset, we use the standard cross entropy as the training objective function:
\begin{equation}
\mathcal{L}_{1}=\mathbb{E}_{\mathcal{D}_{\mathrm{train}}}\left[\ell\left(f(\boldsymbol{x} ; \boldsymbol{\theta}), y\right)\right]=\mathbb{E}_{\mathcal{D}_{\mathrm{train}}}\left[-\boldsymbol{e}^{y} \log f\left(\boldsymbol{x} ; \boldsymbol{\theta}\right)\right],
\end{equation}
For the auxiliary dataset, we enforce the outputs to be consistent with dynamic noise labels, uniformly sampled from the label set in each epoch. Therefore, the training objective function is as follows:
\begin{equation}
\label{eq:aux_loss}
\mathcal{L}_{2}=\mathbb{E}_{\mathcal{D}_{\mathrm{out}}}\left[\ell\left(f(\widetilde{\boldsymbol{x}} ; \boldsymbol{\theta}), \widetilde{y}\right)\right]= \mathbb{E}_{\mathcal{D}_{\mathrm{out}}}\left[-\boldsymbol{e}^{\widetilde{y}} \log f\left(\widetilde{\boldsymbol{x}} ; \boldsymbol{\theta}\right)\right], \text{where}\ \widetilde{y} \sim \mathcal{U}_k
\end{equation}
\noindent where $\mathcal{U}_k$ represents a discrete uniform distribution on the label set $\{1,\ldots, k\}$.
Combining the original training dataset and the auxiliary dataset in the training, now we can formalize ODNL as minimizing the final objective:
\begin{equation}
\label{eq:final_loss}
\mathcal{L}_{\mathrm{total}} = \mathcal{L}_{1} + \eta \cdot \mathcal{L}_{2} = \mathbb{E}_{\mathcal{D}_{\mathrm{train}}}\left[\ell\left(f(\boldsymbol{x} ; \boldsymbol{\theta}), y\right)\right] + \eta \cdot \mathbb{E}_{\mathcal{D}_{\mathrm{out}}}\left[\ell\left(f(\widetilde{\boldsymbol{x}} ; \boldsymbol{\theta}), \widetilde{y}\right)\right]
\end{equation}
\noindent where $\eta$ denotes the trade-off hyperparameter. The details of ODNL are provided in Algorithm \ref{alg:ODNL}.
% \noindent where $n$ and $m$ denote the batch sizes of original dataset and open-set noisy dataset respectively.
\noindent\textbf{Extensions to Other Robust Training Algorithms.} It is worthy to note that our regularization is a general method and can be easily incorporated into existing robust training algorithms, including sample selection\cite{han2018co, wei2020combating}, loss correction \cite{patrini2017making}, robust loss function \cite{zhang2018generalized,wang2019symmetric}, etc.
% It can be implemented by adding the cross entropy loss as an additional regularization term.
Given the original learning objective $\mathcal{L}_{\mathrm{robust}}$ of the robust methods and the hyperparameter $\eta$, we can formalize the final objective as:
\begin{align}
\mathcal{L}_{\mathrm{total}} &= \mathcal{L}_{\mathrm{robust}} + \eta \cdot \mathcal{L}_{2}
% &= \mathcal{L}_{robust} -\frac{\eta}{m} \cdot \sum_{i=1}^{m} \boldsymbol{e}^{\widetilde{y}_i} \log f\left(\widetilde{\boldsymbol{x}}_{i} ; \boldsymbol{\theta}\right)
\end{align}
\begin{algorithm}[ht]
\caption{Open-set Regularization with Dynamic Noisy Labels}
\label{alg:ODNL}
\begin{algorithmic}[1]
\REQUIRE Training dataset $\mathcal{D}_{\mathrm{train}}$. Open-set auxiliary dataset $\mathcal{D}_{\mathrm{out}}$;\\
\FOR{each iteration}
\STATE Sample a mini-batch of original training data $\{(\boldsymbol{x}_i, y_i)\}^{n}_{i=0}$ from $\mathcal{D}_{\mathrm{train}}$;\\
\STATE Sample a mini-batch of open-set data $\{\widetilde{\boldsymbol{x}}_i\}^{m}_{i=0}$ from $\mathcal{D}_{\mathrm{out}}$;\\
\STATE Generate random noisy label $\widetilde{y}_i \sim \mathcal{U}_k$ for each open-set data $\widetilde{\boldsymbol{x}}_i$;\\
\STATE Perform gradient descent on $f$ with $\mathcal{L}_{\mathrm{total}}$ from Equation (\ref{eq:final_loss});\\
% \FOR{each $(\boldsymbol{x},y) \in \{(\boldsymbol{x}_i, y_i)\}^{n}_{i=0} \cup \{(\widetilde{\boldsymbol{x}}_i, \widetilde{y}_i)\}^{m}_{i=0}$}
% \STATE Perform gradient descent on $f$ on loss $\mathcal{L}(\boldsymbol{x},y)$;\
% \ENDFOR
% \FOR{each $\{(\boldsymbol{x},y), (\widetilde{\boldsymbol{x}},\widetilde{y})\} \in \{(\boldsymbol{x}_i, y_i),(\widetilde{\boldsymbol{x}}_i, \widetilde{y}_i)\}^{n}_{i=0}$}
% \STATE Perform gradient descent on $f$ on loss $\mathcal{L}(\boldsymbol{x},y) + \lambda \mathcal{L}(\widetilde{\boldsymbol{x}},\widetilde{y})$;\
% \ENDFOR
\ENDFOR
\label{code:open_reg}
\end{algorithmic}
\end{algorithm}
\section{Understanding from SGD noise} \label{understanding}
To further understand the effect of our regularization, we provide an analysis from the lens of SGD noise. Following SLN \cite{chen2021noise}, here we consider an original training sample $\left({\boldsymbol{x}},y\right)$ and an open-set sample $\widetilde{\boldsymbol{x}}$ at each SGD step for convenience. In parameter updates, our regularization adds noises to the original gradients with the open-set noisy sample $\widetilde{\boldsymbol{x}}$ and dynamic noisy label $\widetilde{y} \sim \mathcal{U}_{k}$:
\begin{equation}
\label{eq:gradient}
\widetilde{\nabla}_{\boldsymbol{\theta}}\ell_{\mathrm{total}} = \nabla_{\boldsymbol{\theta}}\ell(f(\boldsymbol{x};\boldsymbol{\theta}),y) + \eta \cdot \nabla_{\boldsymbol{\theta}}\ell(f(\widetilde{\boldsymbol{x}};\boldsymbol{\theta}),\widetilde{y})
\end{equation}
\noindent where $\ell$ denotes the cross entropy loss for single sample.
% Since the SGD noise of ODNL is applied at batch level, we consider each batch $\left(x,y\right)$ and $\left(\widetilde{x},\widetilde{y}\right)$ independently in the following propositions.
For convenience, we omit the hyperparameter $\eta$ and use $\boldsymbol{f}, \boldsymbol{\widetilde{f}}$ to indicate the model output $f\left(\boldsymbol{x}; \boldsymbol{\theta}\right)$ for the original training sample and the model output $f\left(\widetilde{\boldsymbol{x}}; \boldsymbol{\theta}\right)$ for the open-set noisy sample. Following the standard notation of the Jacobian matrix, we have $\nabla_{\boldsymbol{\theta}} \ell \in \mathbb{R}^{1 \times \dot{p}}, \nabla_{\boldsymbol{f}} \ell \in \mathbb{R}^{1 \times k}, \nabla_{\boldsymbol{\theta}} \boldsymbol{f} \in \mathbb{R}^{k \times p}, \nabla_{\boldsymbol{\theta}_{i}} \boldsymbol{f} \in \mathbb{R}^{k}$ and $\nabla_{\boldsymbol{\theta}} \boldsymbol{f}_{i} \in \mathbb{R}^{1 \times p}$. The proofs of the following propositions are presented in Appendix \ref{app:proofs}.
% \begin{prop}
% For the cross-entropy loss, Eq.~(\ref{eq:gradient}) induces noise $\boldsymbol{z}$ on $\nabla_{\boldsymbol{\theta}}\mathcal{L}_{1}(\boldsymbol{x},y)$, s.t., $\boldsymbol{z} \in \mathbb{R}^{p}, \boldsymbol{z}_i\sim\text{Multinoulli}\left\{\boldsymbol{M_i}\right\}$, $\boldsymbol{M} \in \mathbb{R}^{p\times k}$ and $\boldsymbol{M}_{i,j} = \frac{\nabla_{\theta_i}\widetilde{f}_j}{\widetilde{f}_j}$, $\forall i \in \left\{1, \ldots, p\right\}$, $j \in \left\{1, \ldots, k\right\}$, where $\frac{\cdot}{\widetilde{f}_j}$ denotes the element-wise division. Note that the expected value of noise on the $i$-th parameter is $\frac{1}{k}\sum_{j=1}^{k}\frac{\nabla_{\theta_i}\widetilde{f}_j}{\widetilde{f}_j}$.
% \end{prop}
% \begin{prop}
% \label{prop:OND}
% For the cross-entropy loss, Eq.~(\ref{eq:gradient}) induces noise $\boldsymbol{z} = - \frac{\nabla_{\boldsymbol{\theta}}\boldsymbol{\widetilde{f}}_j}{\boldsymbol{\widetilde{f}}_j}$ on $\nabla_{\boldsymbol{\theta}}\ell(\boldsymbol{f},y)$, s.t., $\boldsymbol{z} \in \mathbb{R}^{p}, j \sim \mathcal{U}_k$, where $\frac{\cdot}{\boldsymbol{\widetilde{f}}_j}$ denotes the element-wise division. Note that the expected value of noise on the $i$-th parameter $\boldsymbol{\theta}_i$ is $-\frac{1}{k}\sum_{j=1}^{k}\frac{\nabla_{\boldsymbol{\theta}_i}\boldsymbol{\widetilde{f}}_j}{\boldsymbol{\widetilde{f}}_j}$.
% \end{prop}
\begin{prop}
\label{prop:OND}
For the cross-entropy loss, Eq.~(3) induces noise $\boldsymbol{z} = - \frac{\nabla_{\boldsymbol{\theta}}\boldsymbol{\widetilde{f}}_j}{\boldsymbol{\widetilde{f}}_j}$ on $\nabla_{\boldsymbol{\theta}}\ell(\boldsymbol{f},y)$, s.t., $\boldsymbol{z} \in \mathbb{R}^{p}, j \sim \mathcal{U}_k$, where $\frac{\cdot}{\boldsymbol{\widetilde{f}}_j}$ denotes the element-wise division. Note that the expected value of noise on the $i$-th parameter $\boldsymbol{\theta}_i$ is $-\frac{1}{k}\sum_{j=1}^{k}\frac{\nabla_{\boldsymbol{\theta}_i}\boldsymbol{\widetilde{f}}_j}{\boldsymbol{\widetilde{f}}_j}$.
\end{prop}
\noindent\textbf{The effects of SGD noise.} From Proposition \ref{prop:OND}, we find that our regularization induces SGD noise sampled from a uniform distribution over $\{- \nabla_{\boldsymbol{\theta}}\widetilde{\boldsymbol{f}}_j/\widetilde{\boldsymbol{f}}_j\}_{j=1}^{k}$. The effects of the noises can be analyzed from three aspects:
\begin{itemize}
\item Firstly, our regularization yields SGD noises with random direction, uniformly sampled from the gradients of the open-set noisy sample. In this way, the SGD noises make the training difficult to converge and help escape local minima \cite{bottou2012stochastic,hochreiter1997flat,keskar2016large,kleinberg2018alternative, neyshabur2017exploring}. Without the noises, the model would quickly converge to a local minimum because the optimization of parameters always follows the direction of gradient descent during the training.
\item Secondly, the noises $z$ are independently random to the gradients $\nabla_{\boldsymbol{\theta}}\ell(\boldsymbol{f},y)$ from the original training example as the noises are conducted with the open-set sample $\boldsymbol{\widetilde{x}}$ from $\mathcal{D}_{out}$. If closed-set samples are used here, the generated noises
% would be uniformly sampled from $\left\{- \nabla_{\boldsymbol{\theta}}\boldsymbol{f}_j/\boldsymbol{f}_j\right\}, \forall j \in \left\{1,\ldots, k\right\}$ that
might be conflicted with the original gradients, leading to unstable training and limiting the improvement from noises. This argument is clearly supported by the experimental results in Fig.~\ref{fig:motivation} and Fig.~\ref{fig:open_ablation}.
\item Finally, the expected value of noise is not zero so the noise is biased. In particular, the expected value of noise enforces the model $f$ to produce more conservative predictions on Out-of-Distribution samples. Consequently, this feature
% provides the benefits on model calibration and predictive uncertainty,
would lead to performance improvement on OOD detection tasks as shown in Subsection \ref{OOD_exp}.
\end{itemize}
To further demonstrate the importance of open-set samples in the auxiliary dataset, we conduct experiments with synthetic auxiliary datasets consisting of convex combinations $\boldsymbol{x}^{\prime}$ of open-set samples $\boldsymbol{\widetilde{x}}$ and closed-set samples $\boldsymbol{x}$ with a hyperparameter $\alpha$: $\boldsymbol{x}^{\prime} = (1-\alpha) \cdot \boldsymbol{\widetilde{x}} + \alpha \cdot \boldsymbol{x}$.
% where $\boldsymbol{\widetilde{x}}$ and $\boldsymbol{x}$ are sampled from open-set and closed-set dataset respectively.
Here, we use 300K Random Images \cite{hendrycks2019oe} and CIFAR-5m \cite{ho2020denoising, nakkiran2020deep} as the open-set and closed-set dataset. More training details can be found in Appendix \ref{app:exp_settings}. The results in Fig.~\ref{fig:open_ablation} show that using open-set auxiliary dataset achieves better performance on robustness against inherent noises.
\noindent\textbf{Relations to label randomization methods. } Label randomization is also applied in some existing work in the context of deep learning\cite{chen2021noise, lukasik2020does, xie2016disturblabel}. For example, DisturbLabel \cite{xie2016disturblabel} intentionally generates incorrect labels on a small fraction of training data to improve generalization performance under the settings of clean labels, interpreted as an implicit model ensemble. It can be seen as using a part of training data as a closed-set auxiliary dataset. In such a way,
% we build a connection between our method and DisturbLabel.
DisturbLabel can be interpreted as a special case of our method with a closed-set auxiliary dataset.
%As the method assumes that all the observed labels are perfectly correct, this method is not applicable to the problem of learning with noisy labels, as shown in Fig.~\ref{fig:oe_compare}.
To mitigate the issue of label noise, Stochastic Label Noise (SLN) \cite{chen2021noise} introduces a variant of SGD noise induced by adding dynamic and mean-zero perturbations on the labels of training dataset, while Label Smoothing \cite{lukasik2020does} uses a fixed and biased perturbation on the label. The noise gradients induced by SLN are given as follows:
\begin{align}
\label{eq:sln}
\tilde{\nabla}_{\boldsymbol{\theta}} \ell(\boldsymbol{f}, y)=\nabla_{\boldsymbol{\theta}} \ell\left(\boldsymbol{f}, \boldsymbol{e}^{y}+\sigma_{y} \boldsymbol{z}_{y}\right)
\end{align}
\noindent where $\sigma_y > 0, \boldsymbol{z}_y \in \mathcal{R}^k$, $\boldsymbol{z}_y \sim \mathcal{N}\left(0, \boldsymbol{I}_{k\times k}\right)$. %Here, $\sigma_y z_y$ is the noise on the one-hot label $\boldsymbol{e}^{y}$.
\begin{prop}[Proposition 3 in \citet{chen2021noise}]
\label{prop:SLN}
For the cross-entropy loss, Eq.~(\ref{eq:sln}) in SLN induces noise $\boldsymbol{z} \sim \mathcal{N}(0, \sigma^{2}_{y}M)$ on $\nabla_{\boldsymbol{\theta}}\ell(\boldsymbol{f},y)$, s.t., $\boldsymbol{M} \in \mathbb{R}^{p\times p}$ and $\boldsymbol{M}_{i,j} = \left(\frac{\nabla_{\boldsymbol{\theta}_{i}} \boldsymbol{f}}{\boldsymbol{f}}\right)^{T} \frac{\nabla_{\boldsymbol{\theta}_{j}} \boldsymbol{f}}{\boldsymbol{f}}, \forall i,j \in \left\{1, \ldots, p\right\}$, where $\frac{\cdot}{\boldsymbol{f}_j}$ denotes the element-wise division.
\end{prop}
From Proposition \ref{prop:SLN}, we show that SLN induces SGD noises that obey a Gaussian distribution, where the standard deviation can be adjusted by the hyperparameter $\sigma_{y}$. Therefore, SLN and our ODNL can be considered as different variants of SGD noise, helping the network converge to a flat minimum, which is expected
to provide better generalization performance \cite{bottou2012stochastic,hochreiter1997flat,keskar2016large,kleinberg2018alternative, neyshabur2017exploring}. The effects are explicitly presented in Fig.~\ref{fig:loss_sln} and Fig.~\ref{fig:loss_ours}. Moreover, the results in Fig.~\ref{fig:sln_compare} illustrate that our method could further improve the performance of SLN\footnote{We use the official implementation: \url{https://github.com/chenpf1025/SLN}.} under various types of label noise, showing the complementarity between SLN and our method. %The proof for the Proposition \ref{prop:SLN} is presented in Appendix A.
\begin{figure}
\centering
\begin{subfigure}[b]{0.23\textwidth}
\centering
\includegraphics[width=\textwidth]{img/loss_standard.pdf}
\caption{Standard}
\label{fig:loss_standard}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.23\textwidth}
\centering
\includegraphics[width=\textwidth]{img/loss_OE.pdf}
\caption{OE}
\label{fig:loss_oe}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.23\textwidth}
\centering
\includegraphics[width=\textwidth]{img/loss_sln.pdf}
\caption{SLN}
\label{fig:loss_sln}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.23\textwidth}
\centering
\includegraphics[width=\textwidth]{img/loss_Ours.pdf}
\caption{Ours}
\label{fig:loss_ours}
\end{subfigure}
\caption{Loss landscapes around the local minimum of models trained on CIFAR-10 with symmetric-40\% noisy labels. We compare their sharpness with the same-scale z-axis and show the loss distribution by drawing color bars separately. We show that our method and SLN lead to a flat minimum, while the models trained by Standard and OE converge to a sharp minimum.}
\label{loss_landscape}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{img/open_ablation2.pdf}
\caption{The importance of open set}
\label{fig:open_ablation}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{img/sln_compare.pdf}
\caption{Our method can improve SLN}
\label{fig:sln_compare}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{img/oe_compare.pdf}
\caption{Compare with OE.}
\label{fig:oe_compare}
\end{subfigure}
\caption{(a) Test accuracy (\%) of ODNL ($\eta=1$) on CIFAR-10 with symmetric-40\% inherent noisy labels using different synthetic auxiliary datasets: the larger the $\alpha$ is, the closer the auxiliary sample distribution $\mathcal{D}_{\mathrm{out}}$ is to the origin sample distribution $\mathcal{D}_{\mathrm{train}}$. (b) Comparison of performances on CIFAR-10 with different types of inherent noisy labels using SLN and SLN+Ours. Here we follow the settings of SLN and the training details are provided in Appendix \ref{app:exp_settings}. (c) Comparison of performances on CIFAR10 with different types of inherent noisy labels using OE and Ours.}
\label{fig:analyze}
\end{figure}
\noindent\textbf{Relations to Outlier Exposure. } Outlier Exposure (OE) \cite{hendrycks2019oe} is an effective method designed for OOD detection tasks. Specifically, OE also makes use of a large, unlabeled auxiliary dataset sampled from $\mathcal{D}_{\text{out}}$, and regularizes the softmax probabilities to be a uniform distribution for out-of-distribution data in the auxiliary dataset. In such manner, OE train the model to discover signals and learn heuristics to detect whether a query is sampled from $\mathcal{D}_{\mathrm{train}}$ or $\mathcal{D}_{\mathrm{out}}$. The training objective is to minimize: $\mathbb{E}_{(\boldsymbol{x}, y) \sim \mathcal{D}_{\mathrm{train}}}\left[\ell(f(\boldsymbol{x}; \boldsymbol{\theta}), y)\right]+\lambda \cdot \mathbb{E}_{\boldsymbol{\widetilde{x}} \sim \mathcal{D}_{\mathrm{out}}}\left[ - \frac{1}{k}\cdot\sum_{i=1}^{k}\log f_i\left(\widetilde{\boldsymbol{x}};\boldsymbol{\theta}\right)\right]$.
\begin{prop}
\label{prop:OE}
For the cross-entropy loss, each open-set sample in OE induces bias $\boldsymbol{z} = -\frac{1}{k}\sum_{j=1}^{k}\frac{\nabla_{\boldsymbol{\theta}}\boldsymbol{\widetilde{f}}_j}{\boldsymbol{\widetilde{f}}_j}$ on $\nabla_{\boldsymbol{\theta}}\ell(\boldsymbol{f},y)$, s.t., $\boldsymbol{z} \in \mathbb{R}^{p}$, where $\frac{\cdot}{\boldsymbol{\widetilde{f}}_j}$ denotes the element-wise division.
\end{prop}
Comparing Proposition \ref{prop:OE} with Proposition \ref{prop:OND}, our method should have similar regularization effect to OE on OOD examples, because the expectation of the SGD noises from ODNL is equal to the bias induced by OE. However, OE cannot improve the robustness against noisy labels as it does not induce dynamic noises so that the optimization of parameters always follows the direction of gradient descent. The difference is empirically shown in Fig.~\ref{fig:oe_compare} and we further verify it in Subsection \ref{OOD_exp}.
% \noindent\textbf{How does Open-set Regularization benefit Robustness?} To answer this question, we should first look back on memorization in training. It has been shown that DNNs learn simple patterns first, before memorizing \cite{pmlr-v70-arpit17a}. From the experimental results in , we can conjecture that memorizing open-set noise is more difficult than learning patterns from clean labels, but at least not more difficult than memorizing closed-set noisy labels.
\section{Experiments}
In this section, we verify the effectiveness of our method on three benchmark datasets: CIFAR-10/100 and Clothing1M. We show that ODNL could not only significantly improve the robustness against various types of inherent noisy labels in standard training, but also enhance the performance of existing robust training techniques. Then we perform a sensitivity analysis to validate the effect of $\eta$. Finally, we conduct experiments to evaluate the performance of ODNL on OOD detection task.
\begin{table*}[!t]
\centering
\caption{Average test accuracy (\%) with standard deviation on CIFAR-10 under various types of noisy labels (over 5 trials). The bold indicates the improved results by integrating our regularization. }
\label{tab:cifar10_results}
\renewcommand\arraystretch{0.9}
\resizebox{1.00\textwidth}{!}{
\setlength{\tabcolsep}{3mm}{
\begin{tabular}{cccccc}
\toprule
Method & Sym-20\% & Sym-50\% & Asymmetric & Dependent & Open-set \\
\midrule
Standard& 86.93$\pm$0.49& 69.48$\pm$0.58 & 78.34$\pm$0.96 & 68.39$\pm$0.65 & 88.45$\pm$0.57\\
\textbf{+ ODNL (Ours)}& \textbf{91.06$\pm$0.64}& \textbf{82.50$\pm$0.58} & \textbf{90.00$\pm$0.35} & \textbf{85.37$\pm$0.32} & \textbf{91.47$\pm$0.39}\\
\midrule
Decoupling& 88.31$\pm$0.39& 80.81$\pm$0.85 & 84.85$\pm$0.22 & 82.44$\pm$0.54 & 84.69$\pm$0.46\\
\textbf{+ ODNL (Ours)}& \textbf{89.70 $\pm$0.74}& \textbf{81.48$\pm$0.37} & \textbf{87.36$\pm$0.68} & \textbf{84.11$\pm$0.51} & \textbf{86.22$\pm$0.57}\\
\midrule
F-correction& 84.95$\pm$0.27& 72.11$\pm$0.15 & 84.46$\pm$0.96 & 72.44$\pm$0.58 & 85.54$\pm$0.86\\
\textbf{+ ODNL (Ours)}& \textbf{88.99$\pm$0.23}& \textbf{82.42$\pm$0.29} & \textbf{88.50$\pm$0.15} & \textbf{84.70$\pm$0.39} & \textbf{88.92$\pm$0.65}\\
\midrule
PHuber-CE& 90.92$\pm$0.28 & 74.07$\pm$0.49 & 81.26$\pm$0.65 & 75.07$\pm$1.21 & 88.35$\pm$0.21\\
\textbf{+ ODNL (Ours)}& \textbf{92.05$\pm$0.36}& \textbf{82.63$\pm$0.28} & \textbf{88.24$\pm$0.43} & \textbf{84.11$\pm$0.45} & \textbf{90.85$\pm$0.29}\\
\midrule
Co-teaching& 91.72$\pm$0.35& 86.65$\pm$0.43 & 87.56$\pm$0.36 & 88.64$\pm$0.79 & 89.91$\pm$0.58\\
\textbf{+ ODNL (Ours)}& \textbf{93.07$\pm$0.79}& \textbf{89.79$\pm$0.76} & \textbf{88.48$\pm$0.37} & \textbf{90.99$\pm$0.28} & \textbf{91.88$\pm$0.53}\\
\midrule
JoCoR \footnotesize{($\lambda=0.5$)}& 92.68$\pm$0.33& 88.36$\pm$0.64 & 84.06$\pm$0.37 & 90.11$\pm$0.19 & 88.80$\pm$0.54\\
\textbf{+ ODNL (Ours)}& \textbf{93.85$\pm$0.69}& \textbf{91.17$\pm$0.52} & \textbf{88.10$\pm$0.69} & \textbf{92.09$\pm$0.84} & \textbf{91.73$\pm$0.73}\\
% \midrule
% DivideMix & - & 93.18$\pm$0.51& 92.00$\pm$0.74 & 93.92$\pm$0.64 & - \\
% \textbf{+ ODNL (Ours)}& - & \textbf{93.87$\pm$0.65} & \textbf{93.21$\pm$0.28} & \textbf{94.10$\pm$0.22} & - \\
\bottomrule
% DivideMix& 0& 0 & 0 & 0 & 0\\
% + ODNL& 0& 0 & 0 & 0 & 0\\
% \bottomrule
\end{tabular}
}
}
\end{table*}
% \begin{table*}[ht]
% \footnotesize
% \centering
% \renewcommand\arraystretch{1}
% \caption{ Test accuracy (\%) on CIFAR-10/100 using PHuber-CE and PHuber-CE+Ours. The bold indicates the improved results.}
% \label{tab:phuber}
% % \resizebox{1.00\textwidth}{!}{
% \setlength{\tabcolsep}{3mm}{
% \begin{tabular}{ccccccc}
% \toprule
% Dataset & Method & Sym-20\% & Sym-50\% &Asym & Dependent & Open \\
% \midrule
% \multirow{2}*{CIFAR10} & PHuber-CE & 90.92& 74.07 & 81.26& 75.07& 88.35 \\
% & +ODNL (Ours) & \textbf{91.98} & \textbf{ 83.19} & \textbf{89.18}& \textbf{80.87} & \textbf{91.62}\\
% \midrule
% \multirow{2}*{CIFAR100} & PHuber-CE & 71.27& 61.37 & 48.14& 64.80& - \\
% & +ODNL (Ours) & \textbf{ 72.18} & \textbf{63.02} & \textbf{49.27}& \textbf{67.16} & -\\
% \bottomrule
% \end{tabular}
% }
% % }
% \end{table*}
% array([[0.54, 0.97, 0.3 , 0.47, 0.83],
% [0.22, 0.12, 0.82, 0.85, 0.19],
% [0.53, 0.81, 0.86, 0.99, 0.16],
% [0.04, 0.28, 0.71, 0.1 , 0.41],
% [0.99, 0.1 , 0.12, 0.12, 0.82],
% [0.23, 0.97, 0.21, 0.49, 0.93],
% [0.83, 0.75, 0.39, 0.1 , 0.96],
% [0.44, 0.65, 0.74, 0.83, 0.05],
% [0.92, 0.93, 0.32, 0.46, 0.54],
% [0.92, 0.86, 0.91, 0.94, 0.22],
% [0.46, 0.47, 0.94, 0.71, 0.81],
% [0.11, 0.09, 0.17, 0.18, 0.91],
% [0.74, 0.56, 0.61, 0.51, 0.33],
% [0.19, 0.37, 0.93, 0.83, 0.75],
% [0.6 , 0.81, 0.3 , 0.64, 0.88],
% [0.92, 0.43, 0.78, 0.2 , 0.72]])
%88.506, 89.364, 81.432, 81.97
% 86.93 &89.94 &69.48 &75.91 &78.34 &85.56 &68.39 &76.50& 88.45& 91.12
% 88.31 &89.27 &80.81 &80.33 &84.85 &87.03 &82.44 &84.14 &84.69 &86.34
% 84.95 &87.95 &72.11 &80.50 &84.46 &87.56 &72.44 &81.83 &85.54 &89.67
% 87.24 &87.99 &82.28 &82.36 &77.96 &82.84 &84.80 &85.67 &85.73 &87.45
% 86.69 &91.98 &70.05 &77.99 &78.10 &86.18 &70.36 &76.87 &88.09 &91.62
% 91.72 &92.80 &86.65 &89.64 &87.56 &88.39 &88.64 &91.36 &89.91 &92.60
% 92.68 &93.41 &88.36 &90.40 &84.06 &86.46 &90.11 &91.75 &88.80 &90.97
\subsection{Setups} \label{setups}
We comprehensively verify the utility of ODNL\footnote{The code is published on \url{https://github.com/hongxin001/ODNL}} on different types of label noise, including symmetric noise, asymmetric noise\cite{zhang2018generalized}, instance-dependent noise \cite{chen2020beyond} and open-set noise \cite{wang2018iterative} synthesized on CIFAR-10/100 and real-world noise on Clothing1M \cite{xiao2015learning}. Symmetric noise assumes each label has the same probability of flipping to any other classes. We uniformly flip the label to other classes with an overall probability 20\% and 50\%. Asymmetric noise assumes labels might be only flipped to similar classes \cite{chen2021noise, patrini2017making, zhang2018generalized}. Noisy labels are generated by mapping TRUCK $\rightarrow$ AUTOMOBILE, BIRD $\rightarrow$ AIRPLANE, DEER $\rightarrow$ HORSE, and CAT $\leftrightarrow$ DOG with probability 40\% for CIFAR-10. For CIFAR-100, we flip each class into the next circularly with probability 40\%. Instance-dependent noise assumes the mislabeling probability is dependent on each instance's input features \cite{chen2020beyond, chen2021noise, xia2020parts}. We use the instance-dependent noise from PDN \cite{xia2020parts} with a noisy rate 40\%, where the noise is synthesized based on the DNN prediction error. Open-set noise contains samples that do not belong to the label set considered in the classification task \cite{wang2018iterative}. We generate open-set noises on CIFAR-10 by randomly replacing 40\% of its training images with images from CIFAR-100.
% \begin{itemize}
% \item Symmetric noise assumes each label has the same probability of flipping to any other classes. We uniformly flip the label to other classes with an overall probability 20\% and 50\%.
% \item Asymmetric noise assumes labels might be only flipped to similar classes \cite{chen2021noise, patrini2017making, zhang2018generalized}. Noisy labels are generated by mapping TRUCK $\rightarrow$ AUTOMOBILE, BIRD $\rightarrow$ AIRPLANE, DEER $\rightarrow$ HORSE, and CAT $\leftrightarrow$ DOG with probability 40\% for CIFAR-10. For CIFAR-100, we flip each class into the next circularly with probability 40\%. %Following the common setting in existing work \cite{chen2021noise, patrini2017making, zhang2018generalized},
% \item Instance-dependent noise assumes the mislabeling probability is dependent on each instance's input features \cite{chen2020beyond, chen2021noise, xia2020parts}. We use the instance-dependent noise from PDN \cite{xia2020parts} with a noise rate 40\%, where the noise is synthetized based on the DNN prediction error.
% \item Open-set noise contains samples that do not belong to the label set considered in the classification task \cite{wang2018iterative}. We generate open-set noises on CIFAR-10 by randomly replacing 40\% of its training images with images from CIFAR-100.
% \item For real-world noise, we use the large-scale real-world dataset, Clothing1M, which contains 1M training images with noisy labels.
% \end{itemize}
We perform training with WRN-40-2 \cite{zagoruyko2016wide} on CIFAR-10 and CIFAR-100.
% On the real-world dataset Clothing1M, we use ResNet-50 with ImageNet pretrained weights.
% For the experiments on CIFAR-10/100,
The network is trained for 200 epochs using SGD with a momentum of 0.9, a weight decay of 0.0005. In each iteration, both the batch sizes of the original dataset and the open-set auxiliary dataset are set as 128. We set the initial learning rate as 0.1, and reduce it by a factor of 10 after 80 and 140 epochs. We use 5k noisy samples as the validation to tune the hyperparameter $\eta$ in $\{0.1, 0.5, 1, 2.5, 5\}$, then train the model on the full training set and report the average test accuracy in the last 5 epochs. All experiments are repeated five times with different seeds and more training details can be found in Appendix \ref{app:exp_settings}. We use 300K Random Images \cite{hendrycks2019oe} as the open-set auxiliary dataset and more discussions on the different choices of open-set auxiliary dataset are presented in Appendix \ref{app:results}.
% \begin{itemize}
% \item Decoupling \cite{malach2017decoupling}, which updates the parameters only using instances which have different predictions from two classifiers.
% \item F-correction \cite{patrini2017making}. which corrects the prediction by the label transition matrix. As the setting in F-correction, we first train a standard network to estimate the transition matrix Q.
% \item PHuber-CE \cite{menon2020can}, which uses a composite loss-based gradient clipping, a variation of standard gradient clipping for label noise robustness.
% \item Co-teaching \cite{han2018co}, which trains two networks simultaneously and cross-updates parameters of peer networks.
% \item JoCoR \cite{wei2020combating}, which trains two networks simultaneously and reduces divergence between the two classifiers by explicit regularization.
% \item SLN \cite{chen2021noise}, which adds mean-zero Gaussian noise to the one-hot labels, where the noise is independently sampled for each instance in each training step.
% \item DivideMix \cite{li2020dividemix}, which combines co-training of two models, sample selection based on the loss, label correction/guessing based on semi-supervised learning MixMatch \cite{berthelot2019mixmatch}, and other techniques including regularization, augmenting each image twice.
% \end{itemize}
\subsection{CIFAR-10 and CIFAR-100} \label{cifar10}
\begin{table*}[!t]
\centering
\renewcommand\arraystretch{0.5}
\caption{Average test accuracy (\%) with standard deviation on CIFAR-100 under various types of noisy labels (over 5 trials). The bold indicates the improved results by integrating our regularization.}
\label{tab:cifar100_results}
\resizebox{1.00\textwidth}{!}{
\setlength{\tabcolsep}{4mm}{
\begin{tabular}{ccccc}
\toprule
Method & Sym-20\% & Sym-50\% & Asymmetric & Dependent \\
\midrule
Standard& 64.87$\pm$0.81& 48.55$\pm$0.65 & 48.77$\pm$0.86 & 53.94$\pm$0.52 \\
\textbf{+ ODNL (Ours)}& \textbf{68.82$\pm$0.24}& \textbf{54.08$\pm$0.54} & \textbf{58.61$\pm$0.25} & \textbf{62.45$\pm$0.41}\\
\midrule
Decoupling& 64.26$\pm$0.23& 49.09$\pm$0.81 & 49.92$\pm$0.71 & 55.17$\pm$0.34 \\
\textbf{+ ODNL (Ours)}&\textbf{66.86$\pm$0.35}& \textbf{51.89$\pm$0.47} & \textbf{57.47$\pm$0.49} & \textbf{58.52$\pm$0.28} \\
\midrule
F-correction&62.05$\pm$0.35& 52.27$\pm$0.15 & 54.45$\pm$0.94 & 57.47$\pm$0.26 \\
\textbf{+ ODNL (Ours)}& \textbf{65.53$\pm$0.28}& \textbf{57.12$\pm$0.65} & \textbf{57.89$\pm$0.55} & \textbf{60.95$\pm$0.81}\\
\midrule
PHuber-CE& 71.27$\pm$0.74& 61.37$\pm$0.28 & 48.14$\pm$0.35 & 64.80$\pm$0.32 \\
\textbf{+ ODNL (Ours)}& \textbf{72.27$\pm$0.53}& \textbf{63.33$\pm$0.48} & \textbf{53.63$\pm$0.72} & \textbf{66.54$\pm$0.60}\\
\midrule
Co-teaching& 72.35$\pm$0.60& 65.21$\pm$0.88 & 54.16$\pm$0.17 & 67.85$\pm$0.82 \\
\textbf{+ ODNL (Ours)}& \textbf{73.12$\pm$0.20}& \textbf{65.46$\pm$0.49} & \textbf{54.92$\pm$0.31} & \textbf{68.37$\pm$0.12} \\
\midrule
JoCoR \footnotesize{($\lambda=0.5$)}& 73.44$\pm$0.19& 67.53$\pm$0.79 & 57.02$\pm$0.32 & 70.01$\pm$0.26 \\
\textbf{+ ODNL (Ours)}& \textbf{74.48$\pm$0.16}& \textbf{68.07$\pm$0.47} & \textbf{58.44$\pm$0.34} & \textbf{71.28$\pm$0.31}\\
% \midrule
% SLN& 88.31$\pm$0.81& 88.31$\pm$0.39 & 88.31$\pm$0.63 & 88.31$\pm$0.23 \\
% \textbf{+ ODNL (Ours)}& \textbf{89.94$\pm$0.53}& \textbf{89.94$\pm$0.73} & \textbf{89.94$\pm$0.33} & \textbf{89.94$\pm$0.32}\\
\bottomrule
% DivideMix& 0& 0 & 0 & 0 & 0\\
% + ODNL& 0& 0 & 0 & 0 & 0\\
% \bottomrule
\end{tabular}
}
}
\end{table*}
% \multirow{2}*{CIFAR100} & PHuber-CE & 71.27& 61.37 & 48.14& 64.80& - \\
% & +ODNL (Ours) & \textbf{ 72.18} & \textbf{63.02} & \textbf{49.27}& \textbf{67.16} & -\\
On CIFAR-10 and CIFAR-100, we verify that ODNL can boost existing robust training methods by integrating ODNL with the following methods: 1) Decoupling \cite{malach2017decoupling}, which updates the parameters only using instances that have different predictions from two classifiers. 2) F-correction \cite{patrini2017making}, which corrects the prediction by the label transition matrix. As the setting in F-correction, we first train a standard network to estimate the transition matrix Q. 3) PHuber-CE \cite{menon2020can}, which uses a composite loss-based gradient clipping, a variation of standard gradient clipping for label noise robustness. 4) Co-teaching \cite{han2018co}, which trains two networks simultaneously and cross-updates parameters of peer networks. 5) JoCoR \cite{wei2020combating}, which trains two networks simultaneously and reduces the divergence between the two classifiers by explicit regularization.
% 6) DivideMix \cite{li2020dividemix}, which combines co-training of two models, sample selection based on the loss, label correction/guessing based on semi-supervised learning MixMatch\cite{berthelot2019mixmatch}, and other techniques including regularization, augmenting each image twice.
% \begin{table}
% \begin{minipage}[!t]{0.48\textwidth}
% \renewcommand{\arraystretch}{0.8}
% \caption{Average test accuracy (\%) with standard deviation of DivideMix on CIFAR-10/100 under various types of noisy labels (over 5 trials). All experiments are implemented based on DivideMix's official implementation. The bold indicates the improved results by the proposed regularization.}
% \label{table1}
% \centering
% \resizebox{1.00\textwidth}{!}{
% \setlength{\tabcolsep}{0.5mm}{
% \begin{tabular}{ccc}
% \toprule
% Method & DivideMix & + ODNL (Ours) \\
% \midrule
% Sym-50\%& 94.28$\pm$0.31& 95.02$\pm$0.24 \\
% \midrule
% Asymmetric & NA$\pm$NA& NA$\pm$NA \\
% \midrule
% Dependent & 94.06$\pm$0.39& NA$\pm$NA \\
% \bottomrule
% \end{tabular}}}
% \end{minipage}
% % \\[0.5pt]
% \hspace{10pt}
% \begin{minipage}[!t]{0.48\textwidth}
% \renewcommand{\arraystretch}{0.8}
% \caption{Test accuracy (\%) on Clothing1M with pretrained ResNet-50. All experiments are implemented based on DivideMix's official implementation. The bold indicates the improved results by the proposed regularization.}
% \label{table2}
% \centering
% \resizebox{1.00\textwidth}{!}{
% \setlength{\tabcolsep}{0.5mm}{
% \begin{tabular}{ccc}
% \toprule
% Method & \textit{best} & \textit{last} \\
% \midrule
% Standard & 70.17& 68.23 \\
% Forward & 70.46& 69.34 \\
% Co-Teaching & 72.32& 71.41 \\
% \textbf{ODNL (Ours)}& \textbf{72.47}& \textbf{72.38} \\
% \midrule
% SLN& 88.31$\pm$0.39& 88.31$\pm$0.39 \\
% \textbf{+ ODNL (Ours)}& \textbf{89.94$\pm$0.70}& \textbf{89.94$\pm$0.70} \\
% \midrule
% DivideMix& 88.31$\pm$0.39& 88.31$\pm$0.39 \\
% \textbf{+ ODNL (Ours)}&\textbf{89.94$\pm$0.70}& \textbf{89.94$\pm$0.70} \\
% \bottomrule
% \end{tabular}}}
% \end{minipage}
% \end{table}
% \begin{wraptable}{r}{0.48\linewidth}
% % \footnotesize
% \centering
% \renewcommand\arraystretch{1}
% \caption{Average test accuracy (\%) with standard deviation of DivideMix on CIFAR-10/100 under various types of noisy labels (over 5 trials). All experiments are implemented based on DivideMix's official implementation. The bold indicates the improved results by the proposed regularization.}
% \label{tab:synthesized_results}
% \setlength{\tabcolsep}{2mm}{
% \begin{tabular}{ccc}
% \toprule
% Method & DivideMix & + ODNL (Ours) \\
% \midrule
% Sym-50\%& 94.28$\pm$0.31& 95.02$\pm$0.24 \\
% \midrule
% Asymmetric & NA$\pm$NA& NA$\pm$NA \\
% \midrule
% Dependent & 94.06$\pm$0.39& NA$\pm$NA \\
% \bottomrule
% % DivideMix& 0& 0 & 0 & 0 & 0\\
% % + ODNL& 0& 0 & 0 & 0 & 0\\
% % \bottomrule
% \end{tabular}
% }
% \end{wraptable}
% \begin{table*}[!t]
% \footnotesize
% \centering
% % \renewcommand\arraystretch{1}
% \caption{Test accuracy (\%) on Clothing1M with pretrained ResNet-50. All experiments are implemented based on DivideMix's official implementation. The bold indicates the improved results by the proposed regularization.}
% \label{tab:synthesized_results}
% % \resizebox{1.00\textwidth}{!}{
% % \setlength{\tabcolsep}{2mm}{
% \begin{tabular}{ccccccccc}
% \toprule
% Method & Standard & Forward & Co-Teaching & \textbf{ODNL (Ours)} & SLN & \textbf{+ ODNL (Ours)} & DivideMix & \textbf{+ ODNL (Ours)}\\
% \midrule
% \textit{best} &70.17& 68.23 & 70.17& 68.23 & 70.17& 68.23 & 70.17& 68.23
% \midrule
% \textit{last} & 88.31 & 88.31 & 70.17& 68.23 & 70.17& 68.23 & 70.17& 68.23 \\
% \bottomrule
% % DivideMix& 0& 0 & 0 & 0 & 0\\
% % + ODNL& 0& 0 & 0 & 0 & 0\\
% % \bottomrule
% \end{tabular}
% % }
% % }
% \end{table*}
% \multirow{2}*{CIFAR10} & PHuber-CE & 90.92& 74.07 & 81.26& 75.07& 88.35 \\
% & +ODNL (Ours) & \textbf{91.98} & \textbf{ 83.19} & \textbf{89.18}& \textbf{80.87} & \textbf{91.62}\\
% \midrule
% \multirow{2}*{CIFAR100} & PHuber-CE & 71.27& 61.37 & 48.14& 64.80& - \\
% & +ODNL (Ours) & \textbf{ 72.18} & \textbf{63.02} & \textbf{49.27}& \textbf{67.16} & -\\
% \bottomrule
\begin{wrapfigure}{r}{0.4\textwidth}
\centering
\includegraphics[width=0.4\textwidth]{img/param.pdf}
\caption{Results of sensitivity analysis on CIFAR-10 with different $\eta$.}
\label{fig:eta_analysis}
% \vspace{-10pt}
\end{wrapfigure}
% & 70.17& 68.23 \\
% Forward & 70.46& 69.34 \\
% Co-Teaching & 72.32& 71.41 \\
% \textbf{ODNL (Ours)}& \textbf{72.47}& \textbf{72.38} \\
% \textbf{+ ODNL (Ours)}& \textbf{89.94$\pm$0.70}& \textbf{89.94$\pm$0.70} \\
% \midrule
% DivideMix& 88.31$\pm$0.39& 88.31$\pm$0.39 \\
% \textbf{+ ODNL (Ours)}&\textbf{89.94$\pm$0.70}& \textbf{89.94$\pm$0.70} \\
The average test accuracy with WRN-40-2 \cite{zagoruyko2016wide} are reported in Table \ref{tab:cifar10_results} and Table \ref{tab:cifar100_results}. The results show that incorporating our regularization into existing robust training methods consistently improves their performance against various types of noisy labels. In particular, we can observe that our method can boost almost all types of existing methods, including robust loss function, sample selection, loss correction, etc., showing that our regularization based upon an open-set auxiliary dataset is a complementary method to existing robust algorithms. More intriguingly, while most existing methods do not bring benefits to the robustness against open-set noisy labels in the training set of CIFAR-10, our regularization could still enhance the performance of all existing methods in this case.
To further illustrate the influence of $\eta$, we present a sensitivity analysis on CIFAR-10 with symmetric-40\% noisy labels in Fig.~\ref{fig:eta_analysis}. Specifically, we highlight the differences in the trend of test accuracy after the second decay of learning rate at the 140th epoch. We can observe that with a large $\eta$, the decreasing trend of test accuracy caused by inherent noisy labels is nearly eliminated. The result verifies that our regularization is an effective method to reduce the overfitting of inherent noisy labels.
\begin{table*}[!t]
\footnotesize
\centering
\renewcommand\arraystretch{1}
\caption{ Test accuracy (\%) on Clothing1M with pretrained ResNet-50. All experiments are implemented based on DivideMix's official implementation. The bold indicates the improved results.}
\label{tab:Clothing1m}
\resizebox{1.00\textwidth}{!}{
\setlength{\tabcolsep}{1mm}{
\begin{tabular}{ccccccccc}
\toprule
Method & Standard & Forward & Co-teaching & \textbf{ODNL (Ours)} & SLN & \textbf{+ ODNL (Ours)} & DivideMix & \textbf{+ ODNL (Ours)}\\
\midrule
\textit{best} &70.17& 70.46 & 71.32& \textbf{72.47} & 72.31& \textbf{73.12} & 74.32& \textbf{74.89} \\
\midrule
\textit{last} & 68.23 & 69.34 & 71.02& \textbf{72.38} & 71.94& \textbf{72.98} & 73.83& \textbf{74.35} \\
\bottomrule
\end{tabular}
}}
\end{table*}
% 64.87 68.98 48.55 54.35 48.77 55.07 53.94 60.47 64.63 66.94
% 64.26 67.06 49.09 53.65 49.92 57.74 55.17 59.32 58.64 60.18
% 62.05 65.60 52.27 57.34 54.45 57.60 57.47 61.83 60.40 63.36
% 65.17 72.18 48.46 63.02 47.64 49.27 53.56 67.16 64.00 67.67
% 72.35 73.34 65.21 65.43 54.16 51.02 67.85 68.41 68.92 69.33
% 74.44 74.74 67.53 68.17 57.02 58.57 70.01 70.68 70.23 71.10
\subsection{Clothing1M}
Clothing1M \cite{xiao2015learning} is a large-scale dataset with real-world noisy labels of clothing images. There are 1 million noisy samples for training, 14K and 10K clean examples for validation and test respectively. Following DivideMix \cite{li2020dividemix}, we conduct experiments by sampling a class-balanced training subset in
each epoch and use ResNet-50 with ImageNet pretrained weights. We set the $\eta$ of the vanilla ODNL as 5. For the auxiliary dataset in our method, we use ImageNet 2012 dataset \cite{ILSVRC15} and remove all classes related to clothes. Other training details strictly follow the settings of DivideMix \cite{li2020dividemix}, presented in Appendix \ref{app:exp_settings}. As shown in Table \ref{tab:Clothing1m}, \textit{best} denotes the score of the epoch where the validation accuracy is optimal, and \textit{last} denotes the scores at the last epoch. The vanilla ODNL outperforms many simple baselines, and it can further improve the performance of SLN and DivideMix. The results show that our method is applicable for large-scale real-world scenarios.
\subsection{OOD detection with noisy labels} \label{OOD_exp}
Out-of-distribution (OOD) detection is an essential problem for the deployment of deep learning especially in safety-critical applications \cite{hendrycks2019oe,liu2020energy}. From the analysis in Section \ref{understanding}, we know that our regularization should have a similar regularization effect to OE \cite{hendrycks2019oe}, which enforces the model to give conservative outputs on OOD examples. Here, we conduct experiments on CIFAR-10 with symmetric-40\% noisy labels to verify the advantage of our method on OOD detection tasks even in a weakly supervised setting. We expect the trained models could perform well in the OOD detection task, even if the labels of training dataset are noisy. Experiments on CIFAR-10 with clean labels are also conducted and the results can be found in Appendix \ref{app:results}. The training settings are the same as those in subsection \ref{cifar10}. Following OE \cite{hendrycks2019oe}, we consider the maximum softmax probability (MSP) baseline \cite{hendrycks2016baseline}, which gives the maximum of softmax output as the OOD score. For the OOD test datasets, we use three types of noises and seven common benchmark datasets: Gaussian, Rademacher \cite{hendrycks2019oe}, Blobs \cite{hendrycks2019oe}, Textures \cite{cimpoi2014describing}, SVHN \cite{netzer2011reading}, CIFAR100, Places365 \cite{zhou2017places}, LSUN-Crop \cite{yu2015lsun}, LSUN-Resize \cite{yu2015lsun}, and iSUN \cite{xu2015turkergaze}. We measure the following metrics: (1) the false positive rate (FPR95) of OOD examples when the true positive rate of in-distribution examples is at 95\%; (2) the area under the receiver operating characteristic curve (AUROC); and (3) the area under the precision-recall curve (AUPR). Table \ref{tab:ood} presents the test accuracy and the average performance over the three types of noises and seven OOD test datasets. We can observe that our method achieves impressive improvement on both the test accuracy and the detection performance, while OE does not bring benefit on the generalization performance. The results can further serve as evidence for the analysis in Section \ref{understanding}.
\begin{table*}[!t]
\footnotesize
\centering
\renewcommand\arraystretch{1}
\caption{ OOD detection performance comparison on CIFAR-10 with symmetric-40\% noisy labels. All values are percentages and are averaged over the ten test datasets described in Appendix \ref{app:exp_settings}. $\uparrow$ indicates larger values are better, and $\downarrow$ indicates smaller values are better. Bold numbers are superior results. Detailed results for each OOD test dataset can be found in Appendix \ref{app:results}}
\label{tab:ood}
% \resizebox{1.00\textwidth}{!}{
\setlength{\tabcolsep}{2mm}{
\begin{tabular}{ccccc}
\toprule
Method & Test Accuracy $\uparrow$ & FPR95 $\downarrow$ & AUROC $\uparrow$ & AUPR $\uparrow$ \\
\midrule
MSP & 77.87& 92.01 & 54.64& 19.41\\
MSP+OE & 78.12& 24.91 & 94.24& 78.09 \\
MSP+Ours & \textbf{86.29}& \textbf{13.43} & \textbf{96.67} & \textbf{84.70} \\
\bottomrule
% DivideMix& 0& 0 & 0 & 0 & 0\\
% + ODNL& 0& 0 & 0 & 0 & 0\\
% \bottomrule
\end{tabular}
% }
}
% \vspace{-10pt}
\end{table*}
% 92.01 54.64 19.41
% 26.76 94.02 78.34
% 18.4 95.54 80.66
\section{Related literature}
% \label{appendix:graph}
% This is the graph appendix...
\noindent\textbf{Learning with noisy labels.}
Learning with noisy labels is a vital topic in weakly-supervised learning, attracting much recent interest with several directions.
1) Some methods propose to train on selected samples, using small-loss selection \cite{han2018co, wei2020combating, yu2019does}, GMM distribution \cite{arazo2019unsupervised, li2020dividemix} or (dis)agreement between two models \cite{malach2017decoupling, wei2020combating, yu2019does}.
2) Some methods aim to design sample weighting schemes that give higher weights on clean samples \cite{jiang2017mentornet, liu2015classification, ren2018learning, shu2019meta, wei2020metainfonet}.
3) Loss correction is also a popular direction based on an estimated noise transition matrix \cite{hendrycks2018using, patrini2017making} or the model's predictions \cite{arazo2019unsupervised, chen2020beyond, reed2014training, tanaka2018joint, zheng2020error}.
4) Robust loss functions are also studied to have a theoretical guarantee \cite{feng2020can, ghosh2017robust, lyu2019curriculum, ma2020normalized,wang2019symmetric, xu2019l_dmi, zhang2018generalized}.
5) Some method apply regularization techniques to improve generalization under the settings of label noise \cite{fatras2021wasserstein,hu2019simple, xia2020robust}, like gradient clipping \cite{menon2020can}, label smoothing \cite{lukasik2020does, szegedy2016rethinking}, temporal ensembling \cite{laine2016temporal} and virtual adversarial training \cite{miyato2018virtual}.
6) Some training strategies for combating noisy labels are built based upon semi-supervised learning methods \cite{li2020dividemix, nguyen2019self}. Studying robustness against label noise helps us further understand the essence of training deep neural networks.
\noindent\textbf{Utilizing auxiliary dataset.} To the best of our knowledge, we are the first to explore the benefits of open-set auxiliary dataset in learning with noisy labels. Auxiliary dataset is also utilized in other problem settings of deep learning. For example, pre-training a network on the large ImageNet database \cite{ILSVRC15} can produce general representations that are useful in many fine-tuning applications \cite{zeiler2014visualizing}. Representation learned from images scraped from search engines and photo-sharing websites could improve objective detection performance \cite{chen2015webly, mahajan2018exploring}. OE uses an auxiliary dataset to teach the network better representations for OOD detection \cite{hendrycks2019oe}. To improve robustness against adversarial attack, a popular method is to train on adversarial examples which can be seen as a generated auxiliary dataset \cite{goodfellow2014explaining}. Unlabelled data is also shown to be beneficial for the adversarial robustness \cite{carmon2019unlabeled, uesato2019labels}. Note that the unlabelled data for improving adversarial robustness is required to be in-distribution, our method utilizes out-of-distribution instances to improve robustness against noisy labels. Recently, OAT \cite{lee2021removing} designed to use out-of-distribution data for improving robustness against adversarial examples, a detailed comparison between OAT and our work is provided in Appendix \ref{app:discussion}.
\noindent\textbf{OOD detection.} OOD detection is an essential building block for safely deploying machine learning models in the open world \cite{hendrycks2016baseline, hendrycks2019oe, liu2020energy}. A common baseline for OOD detection is the softmax confidence score \cite{hendrycks2016baseline}. It has been theoretically shown that neural networks with ReLU activation can produce arbitrarily high softmax confidence for OOD inputs \cite{hein2019relu}. To improve the performance, previous methods have explored using artificially synthesized data from GANs \cite{goodfellow2014generative} or unlabeled data \cite{hendrycks2019oe, lee2017training} as auxiliary OOD training data. Energy scores are shown to be better for distinguishing in- and out-of-distribution samples than softmax scores \cite{liu2020energy}. As a side effect, our method could also achieve superior performance in OOD detection tasks even if the labels of training dataset are noisy.
\section{Conclusion and Discussion}
\label{sec:conclusion}
In this paper, we proposed Open-set regularization with Dynamic Noisy Labels (ODNL), a simple yet effective technique that enhances many existing robust training methods to mitigate the issues of inherent noisy labels across various settings. To the best of our knowledge, our method is the first to utilize open-set auxiliary dataset in the problem of noisy labels. We show that ODNL can be considered as a variant of SGD noises that usually leads to a flat minimum. Extensive experiments show that ODNL can help improve the robustness against various types of noisy labels and it is applicable for large-scale real-world scenarios. Moreover, ODNL also brings a ``side effect'': it can improve the OOD detection performance, which is also essential for the deployment of deep learning in safety-critical applications. On the other hand, ODNL is computationally inexpensive and can be easily integrated into existing algorithms.
Overall, our method is an effective and complementary approach for boosting robustness against inherent noisy labels.
% \textbf{Limitations and future directions.} In this work, we propose to utilize open-set auxiliary dataset to improve robustness against noisy labels. From the experiments in Subsection 2.2, we show that open-set noises can be harmless to generalization performance and even benefit the robustness of neural networks against noisy labels. However,
In the future, the observations and analyses in this work can help understand the effect of open-set noises and inspire more specially designed methods using auxiliary dataset to combat label noise in deep learning. There are still some limitations in our current exploration and method as follows.
\textit{instance-dependent open-set noisy labels}. In this work, we only consider open-set noisy labels that are independent from the instances. In real scenarios, an open-set sample with a noisy label might be similar to the clean samples in this class but still do not belong to this class. We may get different observations when we extend the setting to instance-dependent open-set noisy labels.
\textit{Synthesized auxiliary dataset}. In this work, we use realistic open-set dataset in our method. It is also natural to ask whether synthesized data can be used to combating noisy labels. For example, one may generate adversarial examples with dynamic noisy labels to improve robustness against noisy labels.
% \section*{Acknowledgments and Disclosure of Funding}
\begin{ack}
This research was supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG-RP-2019-0013), National Satellite of Excellence in Trustworthy Software Systems (Award No: NSOE-TSS2019-01), and NTU.
\end{ack}
% \section*{Changes for conditional accept}
% According to the meta review, the paper is being conditionally accepted because the previous version of this paper made use of the 80 million tiny images dataset, which has been retracted. Following the NeurIPS ethical guidelines, we replace the 80 Million Tiny Images dataset with the 300K Random Images dataset \footnote{The dataset is published on \url{https://github.com/hendrycks/outlier-exposure}.}, which is a cleaned and debiased dataset provided by the authors of \citet{hendrycks2019oe} from UC Berkeley. In particular, all experiments with 80 Million Tiny Images have been reconducted using the 300K Random Images dataset and all the new results hold the same conclusions as before. The code for reproducing our method with 300K Random Images is also updated.%The updated results include Figure \ref{fig:motivation}, Figure \ref{fig:loss_ours}, Figure \ref{fig:analyze}, Table \ref{tab:cifar10_results}, Table \ref{tab:cifar100_results}, Figure \ref{fig:eta_analysis} and Table \ref{tab:ood}. %Besides, the best values of $\eta$ are reported in Appendix
% \section*{Funding Transparency Statement}
% Funding in direct support of this work: the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG-RP-2019-0013), National Satellite of Excellence in Trustworthy Software Systems (Award No: NSOE-TSS2019-01), and NTU.
% \newpage
{\small
\bibliographystyle{plainnat}
\bibliography{egbib}
}
% \newpage
% \section*{Checklist}
% \begin{enumerate}
% \item For all authors...
% \begin{enumerate}
% \item Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
% \answerYes{}
% \item Did you describe the limitations of your work?
% \answerYes{See Section~\ref{sec:conclusion}.}
% \item Did you discuss any potential negative societal impacts of your work?
% \answerNo{}
% \item Have you read the ethics review guidelines and ensured that your paper conforms to them?
% \answerYes{}
% \end{enumerate}
% \item If you are including theoretical results...
% \begin{enumerate}
% \item Did you state the full set of assumptions of all theoretical results?
% \answerYes{}
% \item Did you include complete proofs of all theoretical results?
% \answerYes{}
% \end{enumerate}
% \item If you ran experiments...
% \begin{enumerate}
% \item Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)?
% \answerYes{}
% \item Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)?
% \answerYes{See Subsection \ref{setups} and Appendix \ref{app:exp_settings}. }
% \item Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)?
% \answerYes{See Fig.~\ref{fig:motivation}. }
% \item Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)?
% \answerYes{See Appendix \ref{app:exp_settings}.}
% \end{enumerate}
% \item If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
% \begin{enumerate}
% \item If your work uses existing assets, did you cite the creators?
% \answerYes{}
% \item Did you mention the license of the assets?
% \answerNA{}{}
% \item Did you include any new assets either in the supplemental material or as a URL?
% \answerNA{}
% \item Did you discuss whether and how consent was obtained from people whose data you're using/curating?
% \answerNA{}
% \item Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content?
% \answerNA{}
% \end{enumerate}
% \item If you used crowdsourcing or conducted research with human subjects...
% \begin{enumerate}
% \item Did you include the full text of instructions given to participants and screenshots, if applicable?
% \answerNA{}
% \item Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable?
% \answerNA{}
% \item Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation?
% \answerNA{}
% \end{enumerate}
% \end{enumerate}
\clearpage
\begin{appendices}
\setcounter{table}{4}
\section{Proofs}
\label{app:proofs}
\begin{prop}
% \label{prop:OND}
For the cross-entropy loss, Eq.~(6) induces noise $\boldsymbol{z} = - \frac{\nabla_{\boldsymbol{\theta}}\boldsymbol{\widetilde{f}}_j}{\boldsymbol{\widetilde{f}}_j}$ on $\nabla_{\boldsymbol{\theta}}\ell(\boldsymbol{f},y)$, s.t., $\boldsymbol{z} \in \mathbb{R}^{p}, j \sim \mathcal{U}_k$, where $\frac{\cdot}{\boldsymbol{\widetilde{f}}_j}$ denotes the element-wise division. Note that the expected value of noise on the $i$-th parameter $\boldsymbol{\theta}_i$ is $-\frac{1}{k}\sum_{j=1}^{k}\frac{\nabla_{\boldsymbol{\theta}_i}\boldsymbol{\widetilde{f}}_j}{\boldsymbol{\widetilde{f}}_j}$.
\end{prop}
\begin{proof}
For Eq.~(6), the noisy gradient is
\begin{align*}
\widetilde{\nabla}_{\boldsymbol{\theta}}\ell_{\mathrm{total}} = \nabla_{\boldsymbol{\theta}}\ell\left(\boldsymbol{f},y\right) + \eta \cdot \nabla_{\boldsymbol{\theta}}\ell\left(\widetilde{\boldsymbol{f}},\widetilde{y}\right) &=\nabla_{\boldsymbol{\theta}}\ell\left(\boldsymbol{f},y\right) + \eta \cdot \nabla_{\widetilde{\boldsymbol{f}}}\ell\left(\widetilde{\boldsymbol{f}},\widetilde{y}\right)\cdot\nabla_{\boldsymbol{\theta}}\widetilde{\boldsymbol{f}}\\
% &= \nabla_{\boldsymbol{\theta}}\ell\left(\boldsymbol{f},y\right) - \eta \cdot \left(\frac{e^{\widetilde{y}}}{\widetilde{\boldsymbol{f}}}\right)^{T}\cdot\nabla_{\boldsymbol{\theta}}\widetilde{\boldsymbol{f}}
\end{align*}
For convenience, we omit the hyperparameter $\eta$.
Note that cross-entropy loss is $\ell\left(\boldsymbol{f},y\right) = -\boldsymbol{e}^{y} \log \boldsymbol{f}$, then the noise on $\nabla_{\boldsymbol{\theta}}\ell\left(\boldsymbol{f},y\right)$ is
% $$\boldsymbol{z}=-\left(\frac{e^{\widetilde{y}}}{\widetilde{\boldsymbol{f}}}\right)^{T}\cdot\nabla_{\boldsymbol{\theta}}\widetilde{\boldsymbol{f}}, s.t., \widetilde{y} \sim \mathcal{U}_{k} \Leftrightarrow \boldsymbol{z} = -\frac{\nabla_{\boldsymbol{\theta}}\boldsymbol{\widetilde{f}}_j}{\boldsymbol{\widetilde{f}}_j}, s.t., j \sim \mathcal{U}_k$$
\begin{equation*}
\boldsymbol{z} = \nabla_{\widetilde{\boldsymbol{f}}}\ell\left(\widetilde{\boldsymbol{f}},\widetilde{y}\right)\cdot\nabla_{\boldsymbol{\theta}}\widetilde{\boldsymbol{f}} = -\left(\frac{e^{\widetilde{y}}}{\widetilde{\boldsymbol{f}}}\right)^{T}\cdot\nabla_{\boldsymbol{\theta}}\widetilde{\boldsymbol{f}} = - \sum_{j=1}^{k} \left(e^{\widetilde{y}}_j \cdot \frac{\nabla_{\boldsymbol{\theta_i}}\boldsymbol{\widetilde{f}}_j}{\boldsymbol{\widetilde{f}}_j}\right), \quad \text{s.t. } \widetilde{y} \sim \mathcal{U}_{k}.
\end{equation*}
Since $\boldsymbol{e}^{\widetilde{y}} = (0, \cdots, 1, \cdots, 0)$ is the one-hot vector and only the $y$-th entry is 1, the expression of the noise $\boldsymbol{z}$ can be simplified as:
\begin{equation*}
\boldsymbol{z} = -\frac{\nabla_{\boldsymbol{\theta}}\boldsymbol{\widetilde{f}}_j}{\boldsymbol{\widetilde{f}}_j}, \quad \text{s.t. } j \sim \mathcal{U}_k.
\end{equation*}
% Note that $\widetilde{y} \sim \mathcal{U}_{k}$, let $z_i$ be the $i$-th entry of $\boldsymbol{z}$, we have
% \begin{equation*}
% z_i = - \sum_{j=1}^{k} \frac{\partial\boldsymbol{f}_j}{\partial\boldsymbol{\theta_i}} \frac{e^{\widetilde{y}}_j}{\boldsymbol{f}_j} = - \sum_{j=1}^{k} \left(e^{\widetilde{y}}_j \cdot \frac{\nabla_{\boldsymbol{\theta_i}}\boldsymbol{\widetilde{f}}_j}{\boldsymbol{\widetilde{f}}_j}\right) \Leftrightarrow - \frac{\nabla_{\boldsymbol{\theta_i}}\boldsymbol{\widetilde{f}}_j}{\boldsymbol{\widetilde{f}}_j}, s.t., j \sim \mathcal{U}_k
% \end{equation*}
Let $z_i$ be the $i$-th entry of $\boldsymbol{z}$, we have
\begin{equation*}
z_i = -\frac{\nabla_{\boldsymbol{\theta}_i}\boldsymbol{\widetilde{f}}_j}{\boldsymbol{\widetilde{f}}_j}, \quad \text{s.t. } j \sim \mathcal{U}_k.
\end{equation*}
Hence,
\begin{equation*}
\mathbb{E}\left(z_i\right) = -\frac{1}{k}\sum_{j=1}^{k}\frac{\nabla_{\boldsymbol{\theta}_i}\boldsymbol{\widetilde{f}}_j}{\boldsymbol{\widetilde{f}}_j}
\end{equation*}
\end{proof}
% \begin{prop}[Proposition 3 in \citet{chen2021noise}]
% \label{prop:SLN}
% For the cross-entropy loss, Eq.~(7) in SLN induces noise $\boldsymbol{z} \sim \mathcal{N}(0, \sigma^{2}_{y}M)$ on $\nabla_{\boldsymbol{\theta}}\ell(\boldsymbol{f},y)$, s.t., $\boldsymbol{M} \in \mathbb{R}^{p\times p}$ and $\boldsymbol{M}_{i,j} = \left(\frac{\nabla_{\boldsymbol{\theta}_{i}} \boldsymbol{f}}{\boldsymbol{f}}\right)^{T} \frac{\nabla_{\boldsymbol{\theta}_{j}} \boldsymbol{f}}{\boldsymbol{f}}, \forall i,j \in \left\{1, \ldots, p\right\}$, where $\frac{\cdot}{\boldsymbol{f}_j}$ denotes the element-wise division.
% \end{prop}
% The proof of Proposition 2 can be found in the Appendix A of \citet{chen2021noise}.
% ~\\
\setcounter{prop}{2}
\begin{prop}
% \label{prop:OE}
For the cross-entropy loss, each open-set sample in OE induces bias $\boldsymbol{z} = -\frac{1}{k}\sum_{j=1}^{k}\frac{\nabla_{\boldsymbol{\theta}}\boldsymbol{\widetilde{f}}_j}{\boldsymbol{\widetilde{f}}_j}$ on $\nabla_{\boldsymbol{\theta}}\ell(\boldsymbol{x},y)$, s.t., $\boldsymbol{z} \in \mathbb{R}^{p}$, where $\frac{\cdot}{\boldsymbol{\widetilde{f}}_j}$ denotes the element-wise division.
\end{prop}
\begin{proof}
The regularization item of OE is $ \ell_{\mathrm{OE}} = - \frac{1}{k}\cdot\sum_{i=1}^{k}\log \boldsymbol{\widetilde{f}}_i$.
Then the gradient of $\ell_{\mathrm{OE}}$ w.r.t $\theta$ is
\begin{equation*}
\nabla_{\theta} \ell_{\mathrm{OE}} = \nabla_{\boldsymbol{\widetilde{f}}}\ell_{\mathrm{OE}} \cdot \nabla_{\theta}\boldsymbol{\widetilde{f}} = \nabla_{\boldsymbol{\widetilde{f}}}\left(- \frac{1}{k}\cdot\sum_{j=1}^{k}\log \boldsymbol{\widetilde{f}}_j\right) \cdot \nabla_{\theta}\boldsymbol{\widetilde{f}} = - \frac{1}{k}\cdot\sum_{j=1}^{k}\frac{\nabla_{\theta}\boldsymbol{\widetilde{f}}_j}{\boldsymbol{\widetilde{f}}_j}
\end{equation*}
\end{proof}
\section{More details on experiment setup}
\label{app:exp_settings}
\subsection{Datasets}
\noindent\textbf{Auxiliary datasets}. 300K Random Images \cite{hendrycks2019oe} is a cleaned and debiased dataset with 300K natural images. In this dataset, Images that belong to CIFAR classes from it, images that belong to Places or LSUN classes, and images with divisive metadata are removed so that $\mathcal{D}_{\mathrm{train}}$ and $\mathcal{D}_{\mathrm{out}}$ are disjoint.
% large-scale, diverse dataset of $32 \times 32$ natural images scrapped from the web.
We use the dataset as the open-set auxiliary dataset for experiments with CIFAR-10 and CIFAR-100. In particular,
% we remove all examples of 80 Million Tiny Images that appear in the CIFAR-10/100 datasets,
For experiments with Clothing1M, ImageNet 2012 dataset \cite{ILSVRC15} is used as the open-set auxiliary dataset and we remove all classes related to clothes.
CIFAR-5m \cite{ho2020denoising, nakkiran2020deep} is a dataset of 6 million synthetic CIFAR-10-like images. For experiments in Fig.~1 and Fig.~3a, we build the closed-set dataset by sampling class-balanced subsets from CIFAR-5m.
\noindent\textbf{OOD test datasets}. Following OE \cite{hendrycks2019oe}, we comprehensively evaluate OOD detectors on artificial and real anomalous distributions, including: Gaussian, Rademacher, Blobs, Textures \cite{cimpoi2014describing}, SVHN \cite{yuval2011svhn}, Places365 \cite{zhou2017places}, LSUN-Crop \cite{yu2015lsun}, LSUN-Resize \cite{yu2015lsun}, iSUN \cite{xu2015turkergaze}. For experiments on CIFAR-10, we also use CIFAR-100 as OOD test dataset. \textit{Gaussian} noises have each dimension i.i.d. sampled from an isotropic Gaussian distribution. \textit{Rademacher} noises are images where each dimension is $-1$ or $1$ with equal probability, so each dimension is sampled from a symmetric Rademacher distribution. \textit{Blobs} noises consist in algorithmically generated amorphous shapes with definite edges.
\textit{Textures} \cite{cimpoi2014describing} is a dataset of describable textural images.
\textit{SVHN} dataset \cite{yuval2011svhn} contains 32 × 32 color images of house numbers. There are ten classes comprised of the digits 0-9.
\textit{Places365} \cite{zhou2017places} consists in images for scene recognition rather than object recognition.
\textit{LSUN} \cite{yu2015lsun} is another scene understanding dataset with fewer classes than Places365. Here we use \textit{LSUN-Crop} and \textit{LSUN-Resize} to denote the cropped and resized version of the LSUN dataset respectively.
\textit{iSUN} \cite{xu2015turkergaze} is a large-scale eye tracking dataset, selected from natural scene images of the SUN database \cite{xiao2010sun}.
%shorturl.at/duvO4
\subsection{Setups}
We conduct all the experiments on NVIDIA GeForce RTX 3090, and implement all methods with default parameters by PyTorch \cite{NEURIPS2019_9015}. Following GCE \cite{zhang2018generalized} and SLN \cite{chen2021noise}, we use 5k noisy samples as the validation set to tune the hyperparameters. We then train the model on the full training set and report the average test accuracy over the last 5 epochs.
\noindent\textbf{For experiments in Fig.~1, Fig.~2, Fig.~3a and Fig.~3c}.
We use WRN-40-2 \cite{zagoruyko2016wide} and the network trains for 200 epochs with a dropout rate of 0.3, using SGD with a momentum of 0.9, a weight decay of 0.0005. In each iteration, both the batch sizes of the original dataset and the open-set auxiliary dataset are set as 128. Note that we compare the performance of training with different-sized auxiliary datasets in Fig.~1, the batch size of the auxiliary dataset is fixed as 128 in all cases. We set the initial learning rate as 0.1, and reduce it by a factor of 10 after 80 and 140 epochs. For drawing loss landscapes in Fig.~2, we use the technique from the \textit{loss-landscapes} library \cite{li2018visualizing}. For experiments in Fig.~3c, we use 40\% noise rate in all types of inherent noisy labels.
\noindent\textbf{For experiments in Fig.~3b}. We use the same setting as that of SLN \cite{chen2021noise}. Specifically, we train WRN-28-2 \cite{zagoruyko2016wide} for 300 epochs using SGD with learning rate 0.001, momentum 0.9, weight decay 0.0005 and a batch size of 128. For the hyperparameter $\sigma$ of SLN, we use $\sigma=1$ for symmetric noise and $\sigma=0.5$ otherwise. Here, we use 40\% noise rate in all types of inherent noisy labels, following the official implementation of SLN.
\noindent\textbf{For experiments on Clothing1M}. We use the same setting as that of DivideMix \cite{li2020dividemix}. Specifically, we use ResNet-50 with ImageNet pretrained weights for 80 epochs. The initial learning rate is set as 0.002 and reduced by a factor of 10 after 40 epochs. For each epoch, we sample 1000 mini-batches from the training data while ensuring the labels are balanced.
\noindent\textbf{Method-specific hyperparameters}. The backbone and training settings of the compared methods are not unified in previous paper and we reimplement all methods in the same backbone for a fair comparison. For experiments on CIFAR-10/100, we set $\lambda=0.5$ for JoCoR \cite{wei2020combating}. For experiments on Clothing1M, we use the default hyperparameters for DivideMix: $M=2, T=0.5, \tau=0.5, \lambda_u=0, \alpha=0.5$, and we set the standard deviation of SLN as $\sigma = 0.2$.
\noindent\textbf{Guideline for tuning $\eta$}. We tune the hyperparameter $\eta$ following a guideline that has been widely used in existing literature like SLN \cite{chen2021noise}. Given a new dataset with unknown noise, we suggest quickly searching the best value of $\eta$ (denoted as $\eta_o$) based on the binary search using the validation accuracy throughout training. 1) If we observe a decrease of validation accuracy at the late stage of training, it implies overfitting and $\eta < \eta_o$. 2) Otherwise, we have $\eta \geq \eta_o$. Based on 1) and 2), we can conduct a binary search to quickly find the best value of $\eta$ even if one would like to search $\eta$ in a very detailed range. The best values of $\eta$ for vanilla ODNL are reported in Table \ref{tab:best_eta}.
\begin{table*}[ht]
\footnotesize
\centering
\renewcommand\arraystretch{1}
\caption{ The best test accuracy (\%) and the value of $\eta$ on CIFAR-10/100 using vanilla ODNL.}
\label{tab:best_eta}
% \resizebox{1.00\textwidth}{!}{
\setlength{\tabcolsep}{3mm}{
\begin{tabular}{ccccccc}
\toprule
Dataset & Method & Sym-20\% & Sym-50\% &Asym & Dependent & Open \\
\midrule
\multirow{2}*{CIFAR-10} & Ours & 91.06 & 82.50 & 90.00 & 85.37 & 91.47 \\
& $\eta$ & 2.5 & 2.5 & 3.0 & 3.5 & 2.0\\
\midrule
\multirow{2}*{CIFAR-100} & Ours & 68.82 & 54.08 & 58.61 & 62.45 & 66.95 \\
& $\eta$ & 1.0 & 1.0 & 2.0 & 2.0 & 1.0 \\
\bottomrule
\end{tabular}
}
\vspace{-10pt}
% }
\end{table*}
\section{More empirical results}
\label{app:results}
\subsection{Our method also works on other architectures.}
% \begin{table*}[ht]
% \footnotesize
% \centering
% \renewcommand\arraystretch{1}
% \caption{ Test accuracy (\%) on CIFAR-10 under 40\% noise rate training with ResNet-34. The bold indicates the improved results.}
% \label{tab:resnet}
% % \resizebox{1.00\textwidth}{!}{
% \setlength{\tabcolsep}{3mm}{
% \begin{tabular}{ccccc}
% \toprule
% Method & Sym & Asym & Dependent & Open \\
% \midrule
% Standard &63.64& 76.38 & 67.15& 88.68 \\
% \midrule
% ODNL (Ours) & \textbf{80.29} & \textbf{85.68} & \textbf{76.67} & \textbf{90.23}\\
% \bottomrule
% \end{tabular}
% }
% % }
% \end{table*}
\begin{table*}[ht]
\footnotesize
\centering
\renewcommand\arraystretch{1}
\caption{ Test accuracy (\%) on CIFAR-10 under symmetric-40\% noise rate training with various architectures. The bold indicates the improved results.}
\label{tab:resnet}
% \resizebox{1.00\textwidth}{!}{
\setlength{\tabcolsep}{3mm}{
\begin{tabular}{cccc}
\toprule
Method & WRN-40-2 & ResNet-18 & VGG-11 \\
\midrule
Standard & 77.55 & 58.60 & 58.94 \\
\midrule
ODNL (Ours) & \textbf{86.29} & \textbf{80.88} & \textbf{74.27}\\
\bottomrule
\end{tabular}
}
% }
\end{table*}
Our regularization method improves robustness against label noise for more than just WRN architectures. Table \ref{tab:resnet} shows that ODNL also improves performance under various types of noisy labels with ResNet-18 \cite{he2016deep} and VGG-11 \cite{simonyan2014very}. In particular, we set weight decay as 0.001 for experiments with ResNet-18 and VGG-11.
% \subsection{Our method can improve PHuber-CE method.}
% \begin{table*}[ht]
% \footnotesize
% \centering
% \renewcommand\arraystretch{1}
% \caption{ Test accuracy (\%) on CIFAR-10/100 using PHuber-CE and PHuber-CE+Ours. The bold indicates the improved results.}
% \label{tab:phuber}
% % \resizebox{1.00\textwidth}{!}{
% \setlength{\tabcolsep}{3mm}{
% \begin{tabular}{ccccccc}
% \toprule
% Dataset & Method & Sym-20\% & Sym-50\% &Asym & Dependent & Open \\
% \midrule
% \multirow{2}*{CIFAR10} & PHuber-CE & 90.92& 74.07 & 81.26& 75.07& 88.35 \\
% & +ODNL (Ours) & \textbf{91.98} & \textbf{ 83.19} & \textbf{89.18}& \textbf{80.87} & \textbf{91.62}\\
% \midrule
% \multirow{2}*{CIFAR100} & PHuber-CE & 71.27& 61.37 & 48.14& 64.80& - \\
% & +ODNL (Ours) & \textbf{ 72.18} & \textbf{63.02} & \textbf{49.27}& \textbf{67.16} & -\\
% \bottomrule
% \end{tabular}
% }
% % }
% \end{table*}
% In the experiments of Table 1 and Table 2, we fix the hyperparameter $\tau$ of the PHuber-CE method \cite{menon2020can} so that the results of PHuber-CE cannot reflect its improvements on robustness against label noise, so we fine-tune the hyperparameter and provide the best performance of PHuber-CE. Here, we use $\tau=10$ for CIFAR-10 and $\tau=30$ for CIFAR-100. The new results are shown in Table \ref{tab:phuber} and we observe that our method can indeed improve the performance of PHuber-CE.
% \newpage
\subsection{Different choices of open-set auxiliary datasets.}
% \begin{wrapfigure}{r}{0.4\textwidth}
% \centering
% \includegraphics[width=0.4\textwidth]{img/size_num.pdf}
% \caption{Average test accuracy (\%) on CIFAR-10 using ODNL ($\eta=1$) with different auxiliary dataset sizes.}
% \label{fig:sample_size}
% \end{wrapfigure}
In the experiments in Subsection 2.2, we found that the sample size of auxiliary dataset is important for the generalization performance when the labels of open-set samples are fixed. To relieve this requirement, we introduce dynamic noisy labels to strengthen our method in Section 3. In particular, our method can use only 50,000 examples from 300K Random Images to achieve comparable performance with using all examples from this dataset. Moreover, we show that using simple Gaussian noises as auxiliary dataset can also improve robustness against noisy labels, but do not work as well as real data. For example, in the case of symmetric-40\% noise on CIFAR-10, using ODNL with simple Gaussian noises achieves 80.24\% accuracy while using ODNL with 300K Random Images achieve 82.03\%. In addition to size and realism, we also find that diversity of auxiliary dataset is not important for robustness against noisy labels, while OE \cite{hendrycks2019oe} claims that it is an important factor for OOD detection. We conduct experiments on CIFAR-10 with symmetric-40\% noises and use the subset of CIFAR-100 with different number of classes as auxiliary dataset. The sample size of these subsets are fixed as 5,000. The results of using different subsets are very close, achieving 81.56\% test accuracy.
% \begin{table*}[!t]
% % \footnotesize
% \centering
% \renewcommand\arraystretch{1}
% \caption{ OOD detection performance comparison on CIFAR-10 under symmetric-40\% noisy labels with different OOD test dataset. $\uparrow$ indicates larger values are better and $\downarrow$ indicates smaller values are better.}
% \label{tab:ood_noise}
% % \resizebox{1.00\textwidth}{!}{
% \setlength{\tabcolsep}{5mm}{
% \begin{tabular}{ccccc}
% \toprule
% OOD test dataset & Method & FPR95 $\downarrow$ & AUROC $\uparrow$ & AUPR $\uparrow$ \\
% \midrule
% \multirow{3}*{Gaussian} & MSP & 66.14 & 72.29 & 27.08\\
% &MSP+OE & 4.47 & 97.74 & 77.17 \\
% & MSP+Ours & 1.84 & 98.88 & 86.56 \\
% \midrule
% \multirow{3}*{Rademacher} & MSP & 91.7& 49.44 & 15.26\\
% &MSP+OE & 12.3& 94.57& 62.02 \\
% & MSP+Ours & 3.58& 97.56 & 74.94 \\
% \midrule
% \multirow{3}*{Blob} & MSP & 99.4& 28.05 & 10.92\\
% &MSP+OE & 12.69& 96.9& 85.09 \\
% & MSP+Ours & 4.51& 98.5 & 89.98 \\
% \midrule
% \multirow{3}*{Textures} & MSP & 98.62& 49.26 & 16.89\\
% &MSP+OE & 33.74& 94.44& 84.19 \\
% & MSP+Ours & 19.96& 95.6 & 84.72 \\
% \midrule
% \multirow{3}*{SVHN} & MSP & 95.78& 56.19 & 19.36\\
% &MSP+OE & 22.57& 95.89 & 86.25 \\
% & MSP+Ours & 17.46& 95.71 & 80.98\\
% \midrule
% \multirow{3}*{CIFAR-100} & MSP & 94.92& 56.2 & 19.86\\
% &MSP+OE & 76.57& 81.71& 60.04\\
% & MSP+Ours & 67.9& 85.64 & 64.77 \\
% \midrule
% \multirow{3}*{LSUN-C} & MSP & 92.03& 60.38 & 22.38\\
% &MSP+OE & 13.99& 97.01& 89.32 \\
% & MSP+Ours & 12.99& 96.48 & 83.17 \\
% \midrule
% \multirow{3}*{LSUN-R} & MSP & 93.36& 60.42 & 22\\
% &MSP+OE & 16.55& 96.3& 83.24 \\
% & MSP+Ours & 8.54& 96.97 & 80.94 \\
% \midrule
% \multirow{3}*{iSUN} & MSP & 92.31& 59.75& 21.6\\
% &MSP+OE & 19.2& 95.79& 81.64 \\
% & MSP+Ours & 9.44& 96.89 & 81.3\\
% \midrule
% \multirow{3}*{Places365} & MSP & 95.84& 54.41 & 18.72\\
% &MSP+OE & 55.5& 89.87 & 74.4 \\
% & MSP+Ours & 37.84& 93.17 & 79.2 \\
% \bottomrule
% \end{tabular}
% % }
% }
% \vspace{-10pt}
% \end{table*}
\begin{table*}[!t]
% \footnotesize
\centering
\renewcommand\arraystretch{1}
\caption{ OOD detection performance comparison on CIFAR-10 under symmetric-40\% noisy labels with different OOD test dataset. $\uparrow$ indicates larger values are better and $\downarrow$ indicates smaller values are better.}
\label{tab:ood_noise}
% \resizebox{1.00\textwidth}{!}{
\setlength{\tabcolsep}{5mm}{
\begin{tabular}{ccccc}
\toprule
OOD test dataset & Method & FPR95 $\downarrow$ & AUROC $\uparrow$ & AUPR $\uparrow$ \\
\midrule
\multirow{3}*{Gaussian} & MSP & 66.14 & 72.29 & 27.08\\
&MSP+OE & 6.58 & 97.39 & 78.37 \\
& MSP+Ours & 2.68 & 98.54 & 83.87 \\
\midrule
\multirow{3}*{Rademacher} & MSP & 91.7& 49.44 & 15.26\\
&MSP+OE & 5.72 & 97.43 & 76.51 \\
& MSP+Ours & 1.19 & 99.34 & 91.30 \\
\midrule
\multirow{3}*{Blob} & MSP & 99.4& 28.05 & 10.92\\
&MSP+OE & 12.51& 97.42& 90.62 \\
& MSP+Ours & 3.46 & 99.13 & 94.81 \\
\midrule
\multirow{3}*{Textures} & MSP & 98.62& 49.26 & 16.89\\
&MSP+OE & 25.43& 94.66 & 79.70 \\
& MSP+Ours & 11.64 & 97.43 & 88.67 \\
\midrule
\multirow{3}*{SVHN} & MSP & 95.78& 56.19 & 19.36\\
&MSP+OE & 44.26& 90.67 & 69.91 \\
& MSP+Ours & 18.53& 95.38 & 79.63\\
\midrule
\multirow{3}*{CIFAR-100} & MSP & 94.92& 56.2 & 19.86\\
&MSP+OE & 71.09& 83.42& 60.41\\
& MSP+Ours & 50.16 & 88.50 & 66.55 \\
\midrule
\multirow{3}*{LSUN-C} & MSP & 92.03& 60.38 & 22.38\\
&MSP+OE & 13.92& 96.68& 86.28 \\
& MSP+Ours & 9.47& 97.52 & 87.47 \\
\midrule
\multirow{3}*{LSUN-R} & MSP & 93.36& 60.42 & 22\\
&MSP+OE & 11.84& 96.77& 82.96 \\
& MSP+Ours & 6.16& 98.11 & 87.37 \\
\midrule
\multirow{3}*{iSUN} & MSP & 92.31& 59.75& 21.6\\
&MSP+OE & 12.34& 96.47& 80.78 \\
& MSP+Ours & 6.50& 98.01 & 86.53\\
\midrule
\multirow{3}*{Places365} & MSP & 95.84& 54.41 & 18.72\\
&MSP+OE & 45.43& 91.54 & 75.38 \\
& MSP+Ours & 24.49& 94.75 & 80.81 \\
\bottomrule
\end{tabular}
% }
}
\vspace{-10pt}
\end{table*}
% \begin{table*}[!t]
% % \footnotesize
% \centering
% \renewcommand\arraystretch{1}
% \caption{ OOD detection performance comparison on CIFAR-10 with clean labels. $\uparrow$ indicates larger values are better and $\downarrow$ indicates smaller values are better.}
% \label{tab:ood_clean}
% % \resizebox{1.00\textwidth}{!}{
% \setlength{\tabcolsep}{5mm}{
% \begin{tabular}{ccccc}
% \toprule
% OOD test dataset & Method & FPR95 $\downarrow$ & AUROC $\uparrow$ & AUPR $\uparrow$ \\
% \midrule
% \multirow{3}*{Gaussian} & MSP&12.45 & 95.71 & 75.55 \\
% &MSP+OE&1.51 & 98.73 & 82.70 \\
% & MSP+Ours&0.75 & 99.56 & 94.21 \\
% \midrule
% \multirow{3}*{Rademacher} & MSP&23.68 & 87.46 & 40.80 \\
% &MSP+OE&2.78 & 97.69 & 74.28 \\
% & MSP+Ours&0.78 & 99.35 & 89.58 \\
% \midrule
% \multirow{3}*{Blob} & MSP&27.09 & 90.79 & 58.76 \\
% &MSP+OE&0.77 & 99.53 & 93.62 \\
% & MSP+Ours&0.97 & 99.36 & 90.93 \\
% \midrule
% \multirow{3}*{Textures} & MSP&48.10 & 87.98 & 59.17 \\
% &MSP+OE&2.80 & 99.16 & 94.86 \\
% & MSP+Ours&3.05 & 99.10 & 95.23 \\
% \midrule
% \multirow{3}*{SVHN} & MSP&20.96 & 92.73 & 65.29 \\
% &MSP+OE&1.27 & 99.51 & 96.72 \\
% & MSP+Ours&2.55 & 99.19 & 94.52 \\
% \midrule
% \multirow{3}*{CIFAR-100} & MSP&48.56 & 87.29 & 55.53 \\
% &MSP+OE&21.35 & 95.28 & 83.17 \\
% & MSP+Ours&23.93 & 94.70 & 81.88 \\
% \midrule
% \multirow{3}*{LSUN-C} & MSP&15.22 & 95.42 & 78.18 \\
% &MSP+OE&2.10 & 99.35 & 96.10 \\
% & MSP+Ours&3.30 & 99.02 & 93.69 \\
% \midrule
% \multirow{3}*{LSUN-R} & MSP&27.31 & 91.43 & 64.92 \\
% &MSP+OE&2.08 & 99.29 & 94.64 \\
% & MSP+Ours&4.31 & 98.72 & 91.56 \\
% \midrule
% \multirow{3}*{iSUN} & MSP&27.36 & 91.48 & 64.65 \\
% &MSP+OE&2.01 & 99.32 & 94.94 \\
% & MSP+Ours&4.02 & 98.81 & 92.25 \\
% \midrule
% \multirow{3}*{Places365} & MSP&50.82 & 87.40 & 57.33 \\
% &MSP+OE&9.61 & 97.69 & 91.78 \\
% & MSP+Ours&11.13 & 97.47 & 90.88 \\
% \hline
% \hline
% \multirow{3}*{Mean} & MSP&30.16 & 90.77 & 62.02 \\
% &MSP+OE&4.63 & 98.56 & 90.28 \\
% & MSP+Ours&5.48 & 98.53 & 91.47\\
% \bottomrule
% \end{tabular}
% % }
% }
% \vspace{-10pt}
% \end{table*}
\begin{table*}[!t]
% \footnotesize
\centering
\renewcommand\arraystretch{1}
\caption{ OOD detection performance comparison on CIFAR-10 with clean labels. $\uparrow$ indicates larger values are better and $\downarrow$ indicates smaller values are better.}
\label{tab:ood_clean}
% \resizebox{1.00\textwidth}{!}{
\setlength{\tabcolsep}{5mm}{
\begin{tabular}{ccccc}
\toprule
OOD test dataset & Method & FPR95 $\downarrow$ & AUROC $\uparrow$ & AUPR $\uparrow$ \\
\midrule
\multirow{3}*{Gaussian} & MSP&12.45 & 95.71 & 75.55 \\
&MSP+OE&0.83 & 99.36 & 90.36 \\
& MSP+Ours&0.74 & 99.59 & 94.87 \\
\midrule
\multirow{3}*{Rademacher} & MSP&23.68 & 87.46 & 40.80 \\
&MSP+OE&0.94 & 97.69 & 87.82 \\
& MSP+Ours&0.56 & 99.73 & 96.95 \\
\midrule
\multirow{3}*{Blob} & MSP&27.09 & 90.79 & 58.76 \\
&MSP+OE&0.96 & 99.43 & 93.08 \\
& MSP+Ours&0.73 & 99.66 & 96.83 \\
\midrule
\multirow{3}*{Textures} & MSP&48.10 & 87.98 & 59.17 \\
&MSP+OE&2.44 & 99.05 & 93.46 \\
& MSP+Ours&3.05 & 99.13 & 95.33 \\
\midrule
\multirow{3}*{SVHN} & MSP&20.96 & 92.73 & 65.29 \\
&MSP+OE&1.60 & 99.20 & 92.58 \\
& MSP+Ours&1.94 & 99.25 & 93.37 \\
\midrule
\multirow{3}*{CIFAR-100} & MSP&48.56 & 87.29 & 55.53 \\
&MSP+OE&21.65 & 94.89 & 81.57 \\
& MSP+Ours&23.53 & 94.56 & 81.26 \\
\midrule
\multirow{3}*{LSUN-C} & MSP&15.22 & 95.42 & 78.18 \\
&MSP+OE&3.51 & 99.15 & 95.15 \\
& MSP+Ours&2.84 & 99.25 & 95.92 \\
\midrule
\multirow{3}*{LSUN-R} & MSP&27.31 & 91.43 & 64.92 \\
&MSP+OE&2.74 & 99.20 & 94.35 \\
& MSP+Ours&3.78 & 98.84 & 92.59 \\
\midrule
\multirow{3}*{iSUN} & MSP&27.36 & 91.48 & 64.65 \\
&MSP+OE&2.71 & 99.17 & 94.30 \\
& MSP+Ours&4.29 & 98.71 & 92.27 \\
\midrule
\multirow{3}*{Places365} & MSP&50.82 & 87.40 & 57.33 \\
&MSP+OE&10.10 & 97.60 & 90.82 \\
& MSP+Ours&10.64 & 97.52 & 90.93 \\
\hline
\hline
\multirow{3}*{Mean} & MSP&30.16 & 90.77 & 62.02 \\
&MSP+OE&4.75 & 98.63 & 91.35 \\
& MSP+Ours&5.21 & 98.62 & 93.03\\
\bottomrule
\end{tabular}
% }
}
\vspace{-10pt}
\end{table*}
\subsection{Detailed results for OOD detection.}
Table \ref{tab:ood_noise} presents the detailed results of OOD detection performance on CIFAR-10 under symmetric-40\% noisy labels with various OOD test datasets. From the results, we show that our method can outperform OE \cite{hendrycks2019oe} in OOD detection tasks when the labels of training dataset are noisy.
Table \ref{tab:ood_clean} presents the detailed results of OOD detection performance on CIFAR-10 under clean labels with various OOD test datasets. We can observe that our method achieve comparable performance to OE \cite{hendrycks2019oe}. The results also support the analysis in Section 4.
%"SVHN", "cifar100", "LSUN-C", "LSUN-R", "iSUN", "Places365"
\section{Discussion}
\label{app:discussion}
\textbf{Relations to OAT.} OAT \cite{lee2021removing} aims to use out-of-distribution data to improve generalization in the context of adversarial robustness. While OAT is relevant to our work, we note that there are key differences in both technique and insight perspectives:
\begin{itemize}
\item Technique: OAT regularizes the softmax probabilities to be a uniform distribution for out-of-distribution data, which is the same as Outlier Exposure \cite{hendrycks2019oe}. As analyzed in Section 4, the regularization in OE and OAT cannot improve the robustness against noisy labels, because it does not induce dynamic noises and the optimization of parameters always follows the direction of gradient descent.
\item Insight: OAT shows that OOD data samples share the same undesirable features as those of the in-distribution data so that these samples could also be used to remove the influence of undesirable features in adversarial robustness. In our work, we show that open-set noisy labels could be harmless and even benefit the robustness against inherent noisy labels. We give an intuitive interpretation for the phenomena by “insufficient capacity” and understand the effects of open-set noises from the perspective of SGD noises.
\end{itemize}
% The open-set noises are generated by uniformly sampling from the label set In real scenarios,
\clearpage
% {\small
% \bibliographystyle{plainnat}
% \bibliography{egbib}
% }
\end{appendices}
\end{document}
\begin{thebibliography}{82}
\providecommand{\natexlab}[1]{#1}
\providecommand{\url}[1]{\texttt{#1}}
\expandafter\ifx\csname urlstyle\endcsname\relax
\providecommand{\doi}[1]{doi: #1}\else
\providecommand{\doi}{doi: \begingroup \urlstyle{rm}\Url}\fi
\bibitem[Arazo et~al.(2019)Arazo, Ortego, Albert, O’Connor, and
McGuinness]{arazo2019unsupervised}
Eric Arazo, Diego Ortego, Paul Albert, Noel O’Connor, and Kevin McGuinness.
\newblock Unsupervised label noise modeling and loss correction.
\newblock In \emph{International Conference on Machine Learning}, pages
312--321. PMLR, 2019.
\bibitem[Arpit et~al.(2017)Arpit, Jastrz{\k{e}}bski, Ballas, Krueger, Bengio,
Kanwal, Maharaj, Fischer, Courville, Bengio, et~al.]{pmlr-v70-arpit17a}
Devansh Arpit, Stanis{\l}aw Jastrz{\k{e}}bski, Nicolas Ballas, David Krueger,
Emmanuel Bengio, Maxinder~S Kanwal, Tegan Maharaj, Asja Fischer, Aaron
Courville, Yoshua Bengio, et~al.
\newblock A closer look at memorization in deep networks.
\newblock In \emph{Proceedings of the 34th International Conference on Machine
Learning}, pages 233--242, 2017.
\bibitem[Blum et~al.(2003)Blum, Kalai, and Wasserman]{blum2003noise}
Avrim Blum, Adam Kalai, and Hal Wasserman.
\newblock Noise-tolerant learning, the parity problem, and the statistical
query model.
\newblock \emph{Journal of the ACM}, 50\penalty0 (4):\penalty0 506--519, 2003.
\bibitem[Bottou(2012)]{bottou2012stochastic}
L{\'e}on Bottou.
\newblock Stochastic gradient descent tricks.
\newblock In \emph{Neural networks: Tricks of the trade}, pages 421--436.
Springer, 2012.
\bibitem[Carmon et~al.(2019)Carmon, Raghunathan, Schmidt, Liang, and
Duchi]{carmon2019unlabeled}
Yair Carmon, Aditi Raghunathan, Ludwig Schmidt, Percy Liang, and John~C Duchi.
\newblock Unlabeled data improves adversarial robustness.
\newblock \emph{arXiv preprint arXiv:1905.13736}, 2019.
\bibitem[Chen et~al.(2020)Chen, Ye, Chen, Zhao, and Heng]{chen2020beyond}
Pengfei Chen, Junjie Ye, Guangyong Chen, Jingwei Zhao, and Pheng-Ann Heng.
\newblock Beyond class-conditional assumption: A primary attempt to combat
instance-dependent label noise.
\newblock \emph{arXiv preprint arXiv:2012.05458}, 2020.
\bibitem[Chen et~al.(2021)Chen, Chen, Ye, Zhao, and Heng]{chen2021noise}
Pengfei Chen, Guangyong Chen, Junjie Ye, Jingwei Zhao, and Pheng-Ann Heng.
\newblock Noise against noise: stochastic label noise helps combat inherent
label noise.
\newblock In \emph{International Conference on Learning Representations}, 2021.
\bibitem[Chen and Gupta(2015)]{chen2015webly}
Xinlei Chen and Abhinav Gupta.
\newblock Webly supervised learning of convolutional networks.
\newblock In \emph{Proceedings of the IEEE International Conference on Computer
Vision}, pages 1431--1439, 2015.
\bibitem[Cimpoi et~al.(2014)Cimpoi, Maji, Kokkinos, Mohamed, and
Vedaldi]{cimpoi2014describing}
Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea
Vedaldi.
\newblock Describing textures in the wild.
\newblock In \emph{Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition}, pages 3606--3613, 2014.
\bibitem[Fatras et~al.(2021)Fatras, Damodaran, Lobry, Flamary, Tuia, and
Courty]{fatras2021wasserstein}
Kilian Fatras, Bharath~Bhushan Damodaran, Sylvain Lobry, Remi Flamary, Devis
Tuia, and Nicolas Courty.
\newblock Wasserstein adversarial regularization for learning with label noise.
\newblock \emph{IEEE Transactions on Pattern Analysis and Machine
Intelligence}, 2021.
\bibitem[Feng et~al.(2020)Feng, Shu, Lin, Lv, Li, and An]{feng2020can}
Lei Feng, Senlin Shu, Zhuoyi Lin, Fengmao Lv, Li~Li, and Bo~An.
\newblock Can cross entropy loss be robust to label noise?
\newblock In \emph{International Joint Conference on Artificial Intelligence},
pages 2206--2212, 2020.
\bibitem[Ghosh et~al.(2017)Ghosh, Kumar, and Sastry]{ghosh2017robust}
Aritra Ghosh, Himanshu Kumar, and PS~Sastry.
\newblock Robust loss functions under label noise for deep neural networks.
\newblock In \emph{Proceedings of the AAAI Conference on Artificial
Intelligence}, volume~31, page 1919–1925, 2017.
\bibitem[Goodfellow et~al.(2014{\natexlab{a}})Goodfellow, Pouget-Abadie, Mirza,
Xu, Warde-Farley, Ozair, Courville, and Bengio]{goodfellow2014generative}
Ian~J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley,
Sherjil Ozair, Aaron Courville, and Yoshua Bengio.
\newblock Generative adversarial networks.
\newblock \emph{arXiv preprint arXiv:1406.2661}, 2014{\natexlab{a}}.
\bibitem[Goodfellow et~al.(2014{\natexlab{b}})Goodfellow, Shlens, and
Szegedy]{goodfellow2014explaining}
Ian~J Goodfellow, Jonathon Shlens, and Christian Szegedy.
\newblock Explaining and harnessing adversarial examples.
\newblock \emph{arXiv preprint arXiv:1412.6572}, 2014{\natexlab{b}}.
\bibitem[Han et~al.(2018)Han, Yao, Yu, Niu, Xu, Hu, Tsang, and
Sugiyama]{han2018co}
Bo~Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and
Masashi Sugiyama.
\newblock Co-teaching: Robust training of deep neural networks with extremely
noisy labels.
\newblock In \emph{Advances in Neural Information Processing Systems}, pages
8527--8537, 2018.
\bibitem[He et~al.(2016)He, Zhang, Ren, and Sun]{he2016deep}
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.
\newblock Deep residual learning for image recognition.
\newblock In \emph{Proceedings of the IEEE conference on Computer Vision and
Pattern Recognition}, pages 770--778, 2016.
\bibitem[Hein et~al.(2019)Hein, Andriushchenko, and Bitterwolf]{hein2019relu}
Matthias Hein, Maksym Andriushchenko, and Julian Bitterwolf.
\newblock Why relu networks yield high-confidence predictions far away from the
training data and how to mitigate the problem.
\newblock In \emph{Proceedings of the IEEE/CVF Conference on Computer Vision
and Pattern Recognition}, pages 41--50, 2019.
\bibitem[Hendrycks and Gimpel(2016)]{hendrycks2016baseline}
Dan Hendrycks and Kevin Gimpel.
\newblock A baseline for detecting misclassified and out-of-distribution
examples in neural networks.
\newblock \emph{arXiv preprint arXiv:1610.02136}, 2016.
\bibitem[Hendrycks et~al.(2018)Hendrycks, Mazeika, Wilson, and
Gimpel]{hendrycks2018using}
Dan Hendrycks, Mantas Mazeika, Duncan Wilson, and Kevin Gimpel.
\newblock Using trusted data to train deep networks on labels corrupted by
severe noise.
\newblock \emph{arXiv preprint arXiv:1802.05300}, 2018.
\bibitem[Hendrycks et~al.(2019)Hendrycks, Mazeika, and
Dietterich]{hendrycks2019oe}
Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich.
\newblock Deep anomaly detection with outlier exposure.
\newblock \emph{Proceedings of the International Conference on Learning
Representations}, 2019.
\bibitem[Ho et~al.(2020)Ho, Jain, and Abbeel]{ho2020denoising}
Jonathan Ho, Ajay Jain, and Pieter Abbeel.
\newblock Denoising diffusion probabilistic models.
\newblock \emph{arXiv preprint arXiv:2006.11239}, 2020.
\bibitem[Hochreiter and Schmidhuber(1997)]{hochreiter1997flat}
Sepp Hochreiter and J{\"u}rgen Schmidhuber.
\newblock Flat minima.
\newblock \emph{Neural Computation}, 9\penalty0 (1):\penalty0 1--42, 1997.
\bibitem[Hu et~al.(2019)Hu, Li, and Yu]{hu2019simple}
Wei Hu, Zhiyuan Li, and Dingli Yu.
\newblock Simple and effective regularization methods for training on noisily
labeled data with generalization guarantee.
\newblock \emph{arXiv preprint arXiv:1905.11368}, 2019.
\bibitem[Jiang et~al.(2017)Jiang, Zhou, Leung, Li, and
Fei-Fei]{jiang2017mentornet}
Lu~Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li~Fei-Fei.
\newblock Mentornet: Learning data-driven curriculum for very deep neural
networks on corrupted labels.
\newblock \emph{arXiv preprint arXiv:1712.05055}, 2017.
\bibitem[Keskar et~al.(2016)Keskar, Mudigere, Nocedal, Smelyanskiy, and
Tang]{keskar2016large}
Nitish~Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy,
and Ping Tak~Peter Tang.
\newblock On large-batch training for deep learning: Generalization gap and
sharp minima.
\newblock \emph{arXiv preprint arXiv:1609.04836}, 2016.
\bibitem[Kleinberg et~al.(2018)Kleinberg, Li, and
Yuan]{kleinberg2018alternative}
Bobby Kleinberg, Yuanzhi Li, and Yang Yuan.
\newblock An alternative view: When does sgd escape local minima?
\newblock In \emph{Proceedings of International Conference on Machine
Learning}, pages 2698--2707. PMLR, 2018.
\bibitem[Krizhevsky et~al.(2009)Krizhevsky, Hinton,
et~al.]{krizhevsky2009learning}
Alex Krizhevsky, Geoffrey Hinton, et~al.
\newblock Learning multiple layers of features from tiny images.
\newblock 2009.
\bibitem[Laine and Aila(2016)]{laine2016temporal}
Samuli Laine and Timo Aila.
\newblock Temporal ensembling for semi-supervised learning.
\newblock \emph{arXiv preprint arXiv:1610.02242}, 2016.
\bibitem[Lee et~al.(2017)Lee, Lee, Lee, and Shin]{lee2017training}
Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin.
\newblock Training confidence-calibrated classifiers for detecting
out-of-distribution samples.
\newblock \emph{arXiv preprint arXiv:1711.09325}, 2017.
\bibitem[Lee et~al.(2019)Lee, Yun, Lee, Lee, Li, and Shin]{lee2019robust}
Kimin Lee, Sukmin Yun, Kibok Lee, Honglak Lee, Bo~Li, and Jinwoo Shin.
\newblock Robust inference via generative classifiers for handling noisy
labels.
\newblock In \emph{Proceedings of International Conference on Machine
Learning}, pages 3763--3772. PMLR, 2019.
\bibitem[Lee et~al.(2021)Lee, Park, Lee, Yi, Lee, and Yoon]{lee2021removing}
Saehyung Lee, Changhwa Park, Hyungyu Lee, Jihun Yi, Jonghyun Lee, and Sungroh
Yoon.
\newblock Removing undesirable feature contributions using out-of-distribution
data.
\newblock 2021.
\bibitem[Li et~al.(2018)Li, Xu, Taylor, Studer, and
Goldstein]{li2018visualizing}
Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein.
\newblock Visualizing the loss landscape of neural nets.
\newblock In \emph{Advances in Neural Information Processing Systems}, pages
6391--–6401, 2018.
\bibitem[Li et~al.(2020)Li, Socher, and Hoi]{li2020dividemix}
Junnan Li, Richard Socher, and Steven~CH Hoi.
\newblock Dividemix: Learning with noisy labels as semi-supervised learning.
\newblock \emph{arXiv preprint arXiv:2002.07394}, 2020.
\bibitem[Liu and Tao(2015)]{liu2015classification}
Tongliang Liu and Dacheng Tao.
\newblock Classification with noisy labels by importance reweighting.
\newblock \emph{IEEE Transactions on pattern analysis and machine
intelligence}, 38\penalty0 (3):\penalty0 447--461, 2015.
\bibitem[Liu et~al.(2020)Liu, Wang, Owens, and Li]{liu2020energy}
Weitang Liu, Xiaoyun Wang, John~D Owens, and Yixuan Li.
\newblock Energy-based out-of-distribution detection.
\newblock \emph{arXiv preprint arXiv:2010.03759}, 2020.
\bibitem[Lukasik et~al.(2020)Lukasik, Bhojanapalli, Menon, and
Kumar]{lukasik2020does}
Michal Lukasik, Srinadh Bhojanapalli, Aditya Menon, and Sanjiv Kumar.
\newblock Does label smoothing mitigate label noise?
\newblock In \emph{Proceedings of International Conference on Machine
Learning}, pages 6448--6458. PMLR, 2020.
\bibitem[Lyu and Tsang(2019)]{lyu2019curriculum}
Yueming Lyu and Ivor~W Tsang.
\newblock Curriculum loss: Robust learning and generalization against label
corruption.
\newblock \emph{arXiv preprint arXiv:1905.10045}, 2019.
\bibitem[Ma et~al.(2020)Ma, Huang, Wang, Romano, Erfani, and
Bailey]{ma2020normalized}
Xingjun Ma, Hanxun Huang, Yisen Wang, Simone Romano, Sarah Erfani, and James
Bailey.
\newblock Normalized loss functions for deep learning with noisy labels.
\newblock In \emph{International Conference on Machine Learning}, pages
6543--6553. PMLR, 2020.
\bibitem[Mahajan et~al.(2018)Mahajan, Girshick, Ramanathan, He, Paluri, Li,
Bharambe, and Van Der~Maaten]{mahajan2018exploring}
Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri,
Yixuan Li, Ashwin Bharambe, and Laurens Van Der~Maaten.
\newblock Exploring the limits of weakly supervised pretraining.
\newblock In \emph{Proceedings of the European Conference on Computer Vision},
pages 181--196, 2018.
\bibitem[Malach and Shalev-Shwartz(2017)]{malach2017decoupling}
Eran Malach and Shai Shalev-Shwartz.
\newblock Decoupling" when to update" from" how to update".
\newblock \emph{arXiv preprint arXiv:1706.02613}, 2017.
\bibitem[Menon et~al.(2020)Menon, Rawat, Kumar, and Reddi]{menon2020can}
Aditya~Krishna Menon, Ankit~Singh Rawat, Sanjiv Kumar, and Sashank Reddi.
\newblock Can gradient clipping mitigate label noise?
\newblock In \emph{International Conference on Learning Representations
(ICLR)}, 2020.
\bibitem[Miyato et~al.(2018)Miyato, Maeda, Koyama, and
Ishii]{miyato2018virtual}
Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii.
\newblock Virtual adversarial training: a regularization method for supervised
and semi-supervised learning.
\newblock \emph{IEEE transactions on Pattern Analysis and Machine
Intelligence}, 41\penalty0 (8):\penalty0 1979--1993, 2018.
\bibitem[Nakkiran et~al.(2020)Nakkiran, Neyshabur, and
Sedghi]{nakkiran2020deep}
Preetum Nakkiran, Behnam Neyshabur, and Hanie Sedghi.
\newblock The deep bootstrap: Good online learners are good offline
generalizers.
\newblock \emph{arXiv preprint arXiv:2010.08127}, 2020.
\bibitem[Netzer et~al.(2011{\natexlab{a}})Netzer, Wang, Coates, Bissacco, Wu,
and Ng]{yuval2011svhn}
Yuval Netzer, T.~Wang, A.~Coates, A.~Bissacco, B.~Wu, and A.~Ng.
\newblock Reading digits in natural images with unsupervised feature learning.
\newblock In \emph{NIPS Workshop on Deep Learning and Unsupervised Feature
Learning}, 2011{\natexlab{a}}.
\bibitem[Netzer et~al.(2011{\natexlab{b}})Netzer, Wang, Coates, Bissacco, Wu,
and Ng]{netzer2011reading}
Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo~Wu, and Andrew~Y
Ng.
\newblock Reading digits in natural images with unsupervised feature learning.
\newblock 2011{\natexlab{b}}.
\bibitem[Neyshabur et~al.(2017)Neyshabur, Bhojanapalli, McAllester, and
Srebro]{neyshabur2017exploring}
Behnam Neyshabur, Srinadh Bhojanapalli, David McAllester, and Nathan Srebro.
\newblock Exploring generalization in deep learning.
\newblock \emph{arXiv preprint arXiv:1706.08947}, 2017.
\bibitem[Nguyen et~al.(2019)Nguyen, Mummadi, Ngo, Nguyen, Beggel, and
Brox]{nguyen2019self}
Duc~Tam Nguyen, Chaithanya~Kumar Mummadi, Thi Phuong~Nhung Ngo, Thi Hoai~Phuong
Nguyen, Laura Beggel, and Thomas Brox.
\newblock Self: Learning to filter noisy labels with self-ensembling.
\newblock \emph{arXiv preprint arXiv:1910.01842}, 2019.
\bibitem[Papadopoulos et~al.(2021)Papadopoulos, Rajati, Shaikh, and
Wang]{papadopoulos2021outlier}
Aristotelis-Angelos Papadopoulos, Mohammad~Reza Rajati, Nazim Shaikh, and
Jiamian Wang.
\newblock Outlier exposure with confidence control for out-of-distribution
detection.
\newblock \emph{Neurocomputing}, 441:\penalty0 138--150, 2021.
\bibitem[Paszke et~al.(2019)Paszke, Gross, Massa, Lerer, Bradbury, Chanan,
Killeen, Lin, Gimelshein, Antiga, Desmaison, Kopf, Yang, DeVito, Raison,
Tejani, Chilamkurthy, Steiner, Fang, Bai, and Chintala]{NEURIPS2019_9015}
Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory
Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban
Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan
Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu~Fang, Junjie Bai, and Soumith
Chintala.
\newblock Pytorch: An imperative style, high-performance deep learning library.
\newblock In \emph{Advances in Neural Information Processing Systems}, pages
8024--8035. Curran Associates, Inc., 2019.
\bibitem[Patrini et~al.(2017)Patrini, Rozza, Krishna~Menon, Nock, and
Qu]{patrini2017making}
Giorgio Patrini, Alessandro Rozza, Aditya Krishna~Menon, Richard Nock, and
Lizhen Qu.
\newblock Making deep neural networks robust to label noise: A loss correction
approach.
\newblock In \emph{Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition}, pages 1944--1952, 2017.
\bibitem[Reed et~al.(2014)Reed, Lee, Anguelov, Szegedy, Erhan, and
Rabinovich]{reed2014training}
Scott Reed, Honglak Lee, Dragomir Anguelov, Christian Szegedy, Dumitru Erhan,
and Andrew Rabinovich.
\newblock Training deep neural networks on noisy labels with bootstrapping.
\newblock \emph{arXiv preprint arXiv:1412.6596}, 2014.
\bibitem[Ren et~al.(2018)Ren, Zeng, Yang, and Urtasun]{ren2018learning}
Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun.
\newblock Learning to reweight examples for robust deep learning.
\newblock In \emph{International Conference on Machine Learning}, pages
4334--4343. PMLR, 2018.
\bibitem[Russakovsky et~al.(2015)Russakovsky, Deng, Su, Krause, Satheesh, Ma,
Huang, Karpathy, Khosla, Bernstein, Berg, and Fei-Fei]{ILSVRC15}
Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma,
Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein,
Alexander~C. Berg, and Li~Fei-Fei.
\newblock {ImageNet Large Scale Visual Recognition Challenge}.
\newblock \emph{International Journal of Computer Vision}, 115\penalty0
(3):\penalty0 211--252, 2015.
\bibitem[Sachdeva et~al.(2021)Sachdeva, Cordeiro, Belagiannis, Reid, and
Carneiro]{sachdeva2021evidentialmix}
Ragav Sachdeva, Filipe~R Cordeiro, Vasileios Belagiannis, Ian Reid, and Gustavo
Carneiro.
\newblock Evidentialmix: Learning with combined open-set and closed-set noisy
labels.
\newblock In \emph{Proceedings of the IEEE/CVF Winter Conference on
Applications of Computer Vision}, pages 3607--3615, 2021.
\bibitem[Shu et~al.(2019)Shu, Xie, Yi, Zhao, Zhou, Xu, and Meng]{shu2019meta}
Jun Shu, Qi~Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng.
\newblock Meta-weight-net: Learning an explicit mapping for sample weighting.
\newblock In \emph{Advances in Neural Information Processing Systems}, pages
1919--1930, 2019.
\bibitem[Simonyan and Zisserman(2014)]{simonyan2014very}
Karen Simonyan and Andrew Zisserman.
\newblock Very deep convolutional networks for large-scale image recognition.
\newblock \emph{arXiv preprint arXiv:1409.1556}, 2014.
\bibitem[Szegedy et~al.(2016)Szegedy, Vanhoucke, Ioffe, Shlens, and
Wojna]{szegedy2016rethinking}
Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew
Wojna.
\newblock Rethinking the inception architecture for computer vision.
\newblock In \emph{Proceedings of the IEEE conference on Computer Vision and
Pattern Recognition}, pages 2818--2826, 2016.
\bibitem[Tanaka et~al.(2018)Tanaka, Ikami, Yamasaki, and
Aizawa]{tanaka2018joint}
Daiki Tanaka, Daiki Ikami, Toshihiko Yamasaki, and Kiyoharu Aizawa.
\newblock Joint optimization framework for learning with noisy labels.
\newblock In \emph{Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition}, pages 5552--5560, 2018.
\bibitem[Uesato et~al.(2019)Uesato, Alayrac, Huang, Stanforth, Fawzi, and
Kohli]{uesato2019labels}
Jonathan Uesato, Jean-Baptiste Alayrac, Po-Sen Huang, Robert Stanforth,
Alhussein Fawzi, and Pushmeet Kohli.
\newblock Are labels required for improving adversarial robustness?
\newblock \emph{arXiv preprint arXiv:1905.13725}, 2019.
\bibitem[Van~Rooyen et~al.(2015)Van~Rooyen, Menon, and
Williamson]{van2015learning}
Brendan Van~Rooyen, Aditya Menon, and Robert~C Williamson.
\newblock Learning with symmetric label noise: The importance of being
unhinged.
\newblock In \emph{Advances in Neural Information Processing Systems}, pages
10--18, 2015.
\bibitem[Wang et~al.(2018)Wang, Liu, Ma, Bailey, Zha, Song, and
Xia]{wang2018iterative}
Yisen Wang, Weiyang Liu, Xingjun Ma, James Bailey, Hongyuan Zha, Le~Song, and
Shu-Tao Xia.
\newblock Iterative learning with open-set noisy labels.
\newblock In \emph{Proceedings of the IEEE conference on Computer Vision and
Pattern Recognition}, pages 8688--8696, 2018.
\bibitem[Wang et~al.(2019)Wang, Ma, Chen, Luo, Yi, and
Bailey]{wang2019symmetric}
Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey.
\newblock Symmetric cross entropy for robust learning with noisy labels.
\newblock In \emph{Proceedings of the IEEE/CVF International Conference on
Computer Vision}, pages 322--330, 2019.
\bibitem[Wei et~al.(2020{\natexlab{a}})Wei, Feng, Chen, and
An]{wei2020combating}
Hongxin Wei, Lei Feng, Xiangyu Chen, and Bo~An.
\newblock Combating noisy labels by agreement: A joint training method with
co-regularization.
\newblock In \emph{Proceedings of the IEEE/CVF Conference on Computer Vision
and Pattern Recognition}, pages 13726--13735, 2020{\natexlab{a}}.
\bibitem[Wei et~al.(2020{\natexlab{b}})Wei, Feng, Wang, and
An]{wei2020metainfonet}
Hongxin Wei, Lei Feng, Rundong Wang, and Bo~An.
\newblock Metainfonet: Learning task-guided information for sample reweighting.
\newblock \emph{arXiv preprint arXiv:2012.05273}, 2020{\natexlab{b}}.
\bibitem[Wu et~al.(2020)Wu, Hu, Xiong, Huan, Braverman, and Zhu]{wu2020noisy}
Jingfeng Wu, Wenqing Hu, Haoyi Xiong, Jun Huan, Vladimir Braverman, and
Zhanxing Zhu.
\newblock On the noisy gradient descent that generalizes as sgd.
\newblock In \emph{International Conference on Machine Learning}, pages
10367--10376. PMLR, 2020.
\bibitem[Xia et~al.(2020{\natexlab{a}})Xia, Liu, Han, Gong, Wang, Ge, and
Chang]{xia2020robust}
Xiaobo Xia, Tongliang Liu, Bo~Han, Chen Gong, Nannan Wang, Zongyuan Ge, and
Yi~Chang.
\newblock Robust early-learning: Hindering the memorization of noisy labels.
\newblock In \emph{International Conference on Learning Representations},
2020{\natexlab{a}}.
\bibitem[Xia et~al.(2020{\natexlab{b}})Xia, Liu, Han, Wang, Deng, Li, and
Mao]{xia2020extended}
Xiaobo Xia, Tongliang Liu, Bo~Han, Nannan Wang, Jiankang Deng, Jiatong Li, and
Yinian Mao.
\newblock Extended t: Learning with mixed closed-set and open-set noisy labels.
\newblock \emph{arXiv preprint arXiv:2012.00932}, 2020{\natexlab{b}}.
\bibitem[Xia et~al.(2020{\natexlab{c}})Xia, Liu, Han, Wang, Gong, Liu, Niu,
Tao, and Sugiyama]{xia2020parts}
Xiaobo Xia, Tongliang Liu, Bo~Han, Nannan Wang, Mingming Gong, Haifeng Liu,
Gang Niu, Dacheng Tao, and Masashi Sugiyama.
\newblock Parts-dependent label noise: Towards instance-dependent label noise.
\newblock \emph{arXiv preprint arXiv:2006.07836}, 2020{\natexlab{c}}.
\bibitem[Xiao et~al.(2010)Xiao, Hays, Ehinger, Oliva, and
Torralba]{xiao2010sun}
Jianxiong Xiao, James Hays, Krista~A Ehinger, Aude Oliva, and Antonio Torralba.
\newblock Sun database: Large-scale scene recognition from abbey to zoo.
\newblock In \emph{proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition}, pages 3485--3492. IEEE, 2010.
\bibitem[Xiao et~al.(2015)Xiao, Xia, Yang, Huang, and Wang]{xiao2015learning}
Tong Xiao, Tian Xia, Yi~Yang, Chang Huang, and Xiaogang Wang.
\newblock Learning from massive noisy labeled data for image classification.
\newblock In \emph{Proceedings of the IEEE conference on Computer Vision and
Pattern Recognition}, pages 2691--2699, 2015.
\bibitem[Xie et~al.(2016)Xie, Wang, Wei, Wang, and Tian]{xie2016disturblabel}
Lingxi Xie, Jingdong Wang, Zhen Wei, Meng Wang, and Qi~Tian.
\newblock Disturblabel: Regularizing cnn on the loss layer.
\newblock In \emph{Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition}, pages 4753--4762, 2016.
\bibitem[Xu et~al.(2015)Xu, Ehinger, Zhang, Finkelstein, Kulkarni, and
Xiao]{xu2015turkergaze}
Pingmei Xu, Krista~A Ehinger, Yinda Zhang, Adam Finkelstein, Sanjeev~R
Kulkarni, and Jianxiong Xiao.
\newblock Turkergaze: Crowdsourcing saliency with webcam based eye tracking.
\newblock \emph{arXiv preprint arXiv:1504.06755}, 2015.
\bibitem[Xu et~al.(2019)Xu, Cao, Kong, and Wang]{xu2019l_dmi}
Yilun Xu, Peng Cao, Yuqing Kong, and Yizhou Wang.
\newblock L\_dmi: A novel information-theoretic loss function for training deep
nets robust to label noise.
\newblock In \emph{Advances in Neural Information Processing Systems}, pages
6222--6233, 2019.
\bibitem[Yan et~al.(2014)Yan, Rosales, Fung, Subramanian, and
Dy]{yan2014learning}
Yan Yan, R{\'o}mer Rosales, Glenn Fung, Ramanathan Subramanian, and Jennifer
Dy.
\newblock Learning from multiple annotators with varying expertise.
\newblock \emph{Machine Learning}, 95\penalty0 (3):\penalty0 291--327, 2014.
\bibitem[Yu et~al.(2015)Yu, Seff, Zhang, Song, Funkhouser, and
Xiao]{yu2015lsun}
Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong
Xiao.
\newblock Lsun: Construction of a large-scale image dataset using deep learning
with humans in the loop.
\newblock \emph{arXiv preprint arXiv:1506.03365}, 2015.
\bibitem[Yu et~al.(2019)Yu, Han, Yao, Niu, Tsang, and Sugiyama]{yu2019does}
Xingrui Yu, Bo~Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama.
\newblock How does disagreement help generalization against label corruption?
\newblock In \emph{International Conference on Machine Learning}, pages
7164--7173. PMLR, 2019.
\bibitem[Zagoruyko and Komodakis(2016)]{zagoruyko2016wide}
Sergey Zagoruyko and Nikos Komodakis.
\newblock Wide residual networks.
\newblock \emph{arXiv preprint arXiv:1605.07146}, 2016.
\bibitem[Zeiler and Fergus(2014)]{zeiler2014visualizing}
Matthew~D Zeiler and Rob Fergus.
\newblock Visualizing and understanding convolutional networks.
\newblock In \emph{European Conference on Computer Vision}, pages 818--833.
Springer, 2014.
\bibitem[Zhang et~al.(2016)Zhang, Bengio, Hardt, Recht, and
Vinyals]{zhang2016understanding}
Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals.
\newblock Understanding deep learning requires rethinking generalization.
\newblock In \emph{Proceedings of International Conference on Learning
Representations}, 2016.
\bibitem[Zhang and Sabuncu(2018)]{zhang2018generalized}
Zhilu Zhang and Mert~R Sabuncu.
\newblock Generalized cross entropy loss for training deep neural networks with
noisy labels.
\newblock \emph{arXiv preprint arXiv:1805.07836}, 2018.
\bibitem[Zheng et~al.(2020)Zheng, Wu, Goswami, Goswami, Metaxas, and
Chen]{zheng2020error}
Songzhu Zheng, Pengxiang Wu, Aman Goswami, Mayank Goswami, Dimitris Metaxas,
and Chao Chen.
\newblock Error-bounded correction of noisy labels.
\newblock In \emph{International Conference on Machine Learning}, pages
11447--11457. PMLR, 2020.
\bibitem[Zhou et~al.(2017)Zhou, Lapedriza, Khosla, Oliva, and
Torralba]{zhou2017places}
Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba.
\newblock Places: A 10 million image database for scene recognition.
\newblock \emph{IEEE Transactions on Pattern Analysis and Machine
Intelligence}, 40\penalty0 (6):\penalty0 1452--1464, 2017.
\end{thebibliography}