Data Optimization in Deep Learning: A Survey

Large-scale, high-quality data are considered an essential factor for the successful application of many deep learning techniques. Meanwhile, numerous real-world deep learning tasks still have to contend with the lack of sufficient amounts of high-quality data. Additionally, issues such as model robustness, fairness, and trustworthiness are also closely related to training data. Consequently, a huge number of studies in the existing literature have focused on the data aspect in deep learning tasks. Some typical data optimization techniques include data augmentation, logit perturbation, sample weighting, and data condensation. These techniques usually come from different deep learning divisions and their theoretical inspirations or heuristic motivations may seem unrelated to each other. This study aims to organize a wide range of existing data optimization methodologies for deep learning from the previous literature, and makes the effort to construct a comprehensive taxonomy for them. The constructed taxonomy considers the diversity of split dimensions, and deep sub-taxonomies are constructed for each dimension. On the basis of the taxonomy, connections among the extensive data optimization methods for deep learning are built in terms of four aspects. We probe into rendering several promising and interesting future directions. The constructed taxonomy and the revealed connections will enlighten the better understanding of existing methods and the design of novel data optimization techniques. Furthermore, our aspiration for this survey is to promote data optimization as an independent subdivision of deep learning. A curated, up-to-date list of resources related to data optimization in deep learning is available at \url{https://github.com/YaoRujing/Data-Optimization}.

Rethinking Class Imbalance in Machine Learning

Imbalance learning is a subfield of machine learning that focuses on learning tasks in the presence of class imbalance. Nearly all existing studies refer to class imbalance as a proportion imbalance, where the proportion of training samples in each class is not balanced. The ignorance of the proportion imbalance will result in unfairness between/among classes and poor generalization capability. Previous literature has presented numerous methods for either theoretical/empirical analysis or new methods for imbalance learning. This study presents a new taxonomy of class imbalance in machine learning with a broader scope. Four other types of imbalance, namely, variance, distance, neighborhood, and quality imbalances between/among classes, which may exist in machine learning tasks, are summarized. Two different levels of imbalance including global and local are also presented. Theoretical analysis is used to illustrate the significant impact of the new imbalance types on learning fairness. Moreover, our taxonomy and theoretical conclusions are used to analyze the shortcomings of several classical methods. As an example, we propose a new logit perturbation-based imbalance learning loss when proportion, variance, and distance imbalances exist simultaneously. Several classical losses become the special case of our proposed method. Meta learning is utilized to infer the hyper-parameters related to the three types of imbalance. Experimental results on several benchmark corpora validate the effectiveness of the proposed method.

Implicit Counterfactual Data Augmentation for Deep Neural Networks

Machine-learning models are prone to capturing the spurious correlations between non-causal attributes and classes, with counterfactual data augmentation being a promising direction for breaking these spurious associations. However, explicitly generating counterfactual data is challenging, with the training efficiency declining. Therefore, this study proposes an implicit counterfactual data augmentation (ICDA) method to remove spurious correlations and make stable predictions. Specifically, first, a novel sample-wise augmentation strategy is developed that generates semantically and counterfactually meaningful deep features with distinct augmentation strength for each sample. Second, we derive an easy-to-compute surrogate loss on the augmented feature set when the number of augmented samples becomes infinite. Third, two concrete schemes are proposed, including direct quantification and meta-learning, to derive the key parameters for the robust loss. In addition, ICDA is explained from a regularization aspect, with extensive experiments indicating that our method consistently improves the generalization performance of popular depth networks on multiple typical learning scenarios that require out-of-distribution generalization.

Combining Adversaries with Anti-adversaries in Training

Adversarial training is an effective learning technique to improve the robustness of deep neural networks. In this study, the influence of adversarial training on deep learning models in terms of fairness, robustness, and generalization is theoretically investigated under more general perturbation scope that different samples can have different perturbation directions (the adversarial and anti-adversarial directions) and varied perturbation bounds. Our theoretical explorations suggest that the combination of adversaries and anti-adversaries (samples with anti-adversarial perturbations) in training can be more effective in achieving better fairness between classes and a better tradeoff between robustness and generalization in some typical learning scenarios (e.g., noisy label learning and imbalance learning) compared with standard adversarial training. On the basis of our theoretical findings, a more general learning objective that combines adversaries and anti-adversaries with varied bounds on each training sample is presented. Meta learning is utilized to optimize the combination weights. Experiments on benchmark datasets under different learning scenarios verify our theoretical findings and the effectiveness of the proposed methodology.

Understanding Difficulty-based Sample Weighting with a Universal Difficulty Measure

Sample weighting is widely used in deep learning. A large number of weighting methods essentially utilize the learning difficulty of training samples to calculate their weights. In this study, this scheme is called difficulty-based weighting. Two important issues arise when explaining this scheme. First, a unified difficulty measure that can be theoretically guaranteed for training samples does not exist. The learning difficulties of the samples are determined by multiple factors including noise level, imbalance degree, margin, and uncertainty. Nevertheless, existing measures only consider a single factor or in part, but not in their entirety. Second, a comprehensive theoretical explanation is lacking with respect to demonstrating why difficulty-based weighting schemes are effective in deep learning. In this study, we theoretically prove that the generalization error of a sample can be used as a universal difficulty measure. Furthermore, we provide formal theoretical justifications on the role of difficulty-based weighting for deep learning, consequently revealing its positive influences on both the optimization dynamics and generalization performance of deep models, which is instructive to existing weighting schemes.

Class-Level Logit Perturbation

Features, logits, and labels are the three primary data when a sample passes through a deep neural network. Feature perturbation and label perturbation receive increasing attention in recent years. They have been proven to be useful in various deep learning approaches. For example, (adversarial) feature perturbation can improve the robustness or even generalization capability of learned models. However, limited studies have explicitly explored for the perturbation of logit vectors. This work discusses several existing methods related to class-level logit perturbation. A unified viewpoint between positive/negative data augmentation and loss variations incurred by logit perturbation is established. A theoretical analysis is provided to illuminate why class-level logit perturbation is useful. Accordingly, new methodologies are proposed to explicitly learn to perturb logits for both single-label and multi-label classification tasks. Extensive experiments on benchmark image classification data sets and their long-tail versions indicated the competitive performance of our learning method. As it only perturbs on logit, it can be used as a plug-in to fuse with any existing classification algorithms. All the codes are available at https://github.com/limengyang1992/lpl.

Exploring the Learning Difficulty of Data Theory and Measure

As learning difficulty is crucial for machine learning (e.g., difficulty-based weighting learning strategies), previous literature has proposed a number of learning difficulty measures. However, no comprehensive investigation for learning difficulty is available to date, resulting in that nearly all existing measures are heuristically defined without a rigorous theoretical foundation. In addition, there is no formal definition of easy and hard samples even though they are crucial in many studies. This study attempts to conduct a pilot theoretical study for learning difficulty of samples. First, a theoretical definition of learning difficulty is proposed on the basis of the bias-variance trade-off theory on generalization error. Theoretical definitions of easy and hard samples are established on the basis of the proposed definition. A practical measure of learning difficulty is given as well inspired by the formal definition. Second, the properties for learning difficulty-based weighting strategies are explored. Subsequently, several classical weighting methods in machine learning can be well explained on account of explored properties. Third, the proposed measure is evaluated to verify its reasonability and superiority in terms of several main difficulty factors. The comparison in these experiments indicates that the proposed measure significantly outperforms the other measures throughout the experiments.

Tackling the Imbalance for GNNs

Different from deep neural networks for non-graph data classification, graph neural networks (GNNs) leverage the information exchange between nodes (or samples) when representing nodes. The category distribution shows an imbalance or even a highly-skewed trend on nearly all existing benchmark GNN data sets. The imbalanced distribution will cause misclassification of nodes in the minority classes, and even cause the classification performance on the entire data set to decrease. This study explores the effects of the imbalance problem on the performances of GNNs and proposes new methodologies to solve it. First, a node-level index, namely, the label difference index ($LDI$), is defined to quantitatively analyze the relationship between imbalance and misclassification. The less samples in a class, the higher the value of its average $LDI$; the higher the $LDI$ of a sample, the more likely the sample will be misclassified. We define a new loss and propose four new methods based on $LDI$. Experimental results indicate that the classification accuracies of the three among our proposed four new methods are better in both transductive and inductive settings. The $LDI$ can be applied to other GNNs.

Which Samples Should be Learned First: Easy or Hard?

An effective weighting scheme for training samples is essential for learning tasks. Numerous weighting schemes have been proposed. Some schemes take the easy-first mode on samples, whereas some others take the hard-first mode. Naturally, an interesting yet realistic question is raised. Which samples should be learned first given a new learning task, easy or hard? To answer this question, three aspects of research are carried out. First, a high-level unified weighted loss is proposed, providing a more comprehensive view for existing schemes. Theoretical analysis is subsequently conducted and preliminary conclusions are obtained. Second, a flexible weighting scheme is proposed to overcome the defects of existing schemes. The three modes, namely, easy/medium/hard-first, can be flexibly switched in the proposed scheme. Third, a wide range of experiments are conducted to further compare the weighting schemes in different modes. On the basis of these works, reasonable answers are obtained. Factors including prior knowledge and data characteristics determine which samples should be learned first in a learning task.