Traditional knowledge distillation focuses on aligning the student's predicted probabilities with both ground-truth labels and the teacher's predicted probabilities. However, the transition to predicted probabilities from logits would obscure certain indispensable information. To address this issue, it is intuitive to additionally introduce a logit-level loss function as a supplement to the widely used probability-level loss function, for exploiting the latent information of logits. Unfortunately, we empirically find that the amalgamation of the newly introduced logit-level loss and the previous probability-level loss will lead to performance degeneration, even trailing behind the performance of employing either loss in isolation. We attribute this phenomenon to the collapse of the classification head, which is verified by our theoretical analysis based on the neural collapse theory. Specifically, the gradients of the two loss functions exhibit contradictions in the linear classifier yet display no such conflict within the backbone. Drawing from the theoretical analysis, we propose a novel method called dual-head knowledge distillation, which partitions the linear classifier into two classification heads responsible for different losses, thereby preserving the beneficial effects of both losses on the backbone while eliminating adverse influences on the classification head. Extensive experiments validate that our method can effectively exploit the information inside the logits and achieve superior performance against state-of-the-art counterparts

Federated learning (FL) enables multiple clients to collaboratively train machine learning models under the coordination of a central server, while maintaining privacy. However, the server cannot directly monitor the local training processes, leaving room for malicious clients to introduce backdoors into the model. Research has shown that backdoor attacks exploit specific neurons that are activated only by malicious inputs, remaining dormant with clean data. Building on this insight, we propose a novel defense method called Flipping Weight Updates of Low-Activation Input Neurons (FLAIN) to counter backdoor attacks in FL. Specifically, upon the completion of global training, we use an auxiliary dataset to identify low-activation input neurons and iteratively flip their associated weight updates. This flipping process continues while progressively raising the threshold for low-activation neurons, until the model's performance on the auxiliary data begins to degrade significantly. Extensive experiments demonstrate that FLAIN effectively reduces the success rate of backdoor attacks across a variety of scenarios,including Non-IID data distributions and high malicious client ratios (MCR), while maintaining minimal impact on the performance of clean data. The source code is available at: FLAIN.

The goal of semi-supervised multi-label learning (SSMLL) is to improve model performance by leveraging the information of unlabeled data. Recent studies usually adopt the pseudo-labeling strategy to tackle unlabeled data based on the assumption that labeled and unlabeled data share the same distribution. However, in realistic scenarios, unlabeled examples are often collected through cost-effective methods, inevitably introducing out-of-distribution (OOD) data, leading to a significant decline in model performance. In this paper, we propose a safe semi-supervised multi-label learning framework based on the theory of evidential deep learning (EDL), with the goal of achieving robust and effective unlabeled data exploitation. On one hand, we propose the asymmetric beta loss to not only compensate for the lack of robustness in common MLL losses, but also to solve the inherent positive-negative imbalance problem faced by the EDL losses in MLL. On the other hand, to construct a robust SSMLL framework, we adopt a dual-head structure to generate class probabilities and instance uncertainties. The former are used to generate pseudo-labels, while the latter are utilized to filter OOD examples. To avoid the need for threshold estimation, we develop a dual-measurement weighted loss function to safely perform unlabeled training. Extensive experiments on multiple benchmark datasets verify the effectiveness of the proposed method in both OOD detection and SSMLL tasks.

Multi-label image classification relies on a large-scale, well-maintained dataset, which may easily be mislabeled due to various subjective reasons. Existing methods for coping with noise usually focus on improving the model robustness in the case of single-label noise. However, compared with noisy single-label learning, noisy multi-label learning is more practical and challenging. To reduce the negative impact of noisy multi-annotations, we propose a universal approach for noisy multi-label learning (UNM). In UNM, we propose the label-wise embedding network which investigates the semantic alignment between label embeddings and their corresponding output features to learn robust feature representations. Meanwhile, mining the co-occurrence of multi-labels is also added to regularize the noisy network predictions. We cyclically change the fitting status of our label-wise embedding network to distinguish the noisy samples and generate pseudo labels for them. As a result, UNM provides an effective way to exploit the label-wise features and semantic label embeddings in noisy scenarios. To verify the generalizability of our method, we also test our method on Partial Multi-label Learning (PML) and Multi-label Learning with Missing Labels (MLML). Extensive experiments on benchmark datasets including Microsoft COCO, Pascal VOC, and Visual Genome explicitly validate the proposed method.

In partial multi-label learning (PML) problems, each instance is partially annotated with a candidate label set, which consists of multiple relevant labels and some noisy labels. To solve PML problems, existing methods typically try to recover the ground-truth information from partial annotations based on extra assumptions on the data structures. While the assumptions hardly hold in real-world applications, the trained model may not generalize well to varied PML tasks. In this paper, we propose a novel approach for partial multi-label learning with meta disambiguation (PML-MD). Instead of relying on extra assumptions, we try to disambiguate between ground-truth and noisy labels in a meta-learning fashion. On one hand, the multi-label classifier is trained by minimizing a confidence-weighted ranking loss, which distinctively utilizes the supervised information according to the label quality; on the other hand, the confidence for each candidate label is adaptively estimated with its performance on a small validation set. To speed up the optimization, these two procedures are performed alternately with an online approximation strategy. Comprehensive experiments on multiple datasets and varied evaluation metrics validate the effectiveness of the proposed method.

Label distribution learning (LDL) is proposed for solving the label ambiguity problem in recent years, which can be seen as an extension of multi-label learning. To improve the performance of label distribution learning, some existing algorithms exploit label correlations in a global manner that assumes the label correlations are shared by all instances. However, the instances in different groups may share different label correlations, and few label correlations are globally applicable in real-world tasks. In this paper, two novel label distribution learning algorithms are proposed by exploiting label correlations on local samples, which are called GD-LDL-SCL and Adam-LDL-SCL, respectively. To utilize the label correlations on local samples, the influence of local samples is encoded, and a local correlation vector is designed as the additional features for each instance, which is based on the different clustered local samples. Then, the label distribution for an unseen instance can be predicted by exploiting the original features and the additional features simultaneously. Extensive experiments on some real-world data sets validate that our proposed methods can address the label distribution problems effectively and outperform state-of-the-art methods.

Partial multi-label learning (PML) deals with problems where each instance is associated with a candidate label set, which contains multiple relevant labels and some noisy labels. In many real-world scenarios, it is impractical to annotate all examples for a huge-size dataset. Instead, a more common case is that only a small set of the data are annotated with partial labels, while most data are unlabeled. In this paper, we formalize such problems as a new learning framework called Semi-Supervised Partial Multi-label Learning (SSPML). To solve the SSPML problem, a latent label variable is introduced for each example as the low-dimensional embedding of the feature space. On one hand, label variables are recovered by encouraging consistent similarity measurement between the feature space and the label space; on the other hand, the similarities are adaptively updated based on the feedback from the label space. Meanwhile, the multi-label classifier is jointly trained under the supervision of label variables. Extensive experiments on multiple datasets from various real-world tasks validate the effectiveness of the proposed approach.

Class-conditional variants of Generative adversarial networks (GANs) have recently achieved a great success due to its ability of selectively generating samples for given classes, as well as improving generation quality. However, its training requires a large set of class-labeled data, which is often expensive and difficult to collect in practice. In this paper, we propose an active sampling method to reduce the labeling cost for effectively training the class-conditional GANs. On one hand, the most useful examples are selected for external human labeling to jointly reduce the difficulty of model learning and alleviate the missing of adversarial training; on the other hand, fake examples are actively sampled for internal model retraining to enhance the adversarial training between the discriminator and generator. By incorporating the two strategies into a unified framework, we provide a cost-effective approach to train class-conditional GANs, which achieves higher generation quality with less training examples. Experiments on multiple datasets, diverse GAN configurations and various metrics demonstrate the effectiveness of our approaches.

Multi-Instance Multi-Label (MIML) learning is a popular framework in machine learning, where each object is represented by a bag of instances, and associated with multiple labels. While MIML learning has achieved success in many applications, it is less clear how the labels are related to the instances. In this paper, we propose to study the causal relation between instances and labels, which on one hand can improve the interpretability of complicated MIML models, and on the other hand may further improve the prediction performance at both instance and bag levels. We exploit prototypes in the instance space as a bridge to represent the examples, and then propose an efficient algorithm to identify the causal relations from prototypes to class labels, which are further utilized for model training and key instance detection. Experiments on various datasets show that in addition to superior classification performance, our approach can identify reasonable causal relations between instances and labels.

Feature missing is a serious problem in many applications, which may lead to low quality of training data and further significantly degrade the learning performance. While feature acquisition usually involves special devices or complex process, it is expensive to acquire all feature values for the whole dataset. On the other hand, features may be correlated with each other, and some values may be recovered from the others. It is thus important to decide which features are most informative for recovering the other features as well as improving the learning performance. In this paper, we try to train an effective classification model with least acquisition cost by jointly performing active feature querying and supervised matrix completion. When completing the feature matrix, a novel objective function is proposed to simultaneously minimize the reconstruction error on observed entries and the supervised loss on training data. When querying the feature value, the most uncertain entry is actively selected based on the variance of previous iterations. In addition, a bi-objective optimization method is presented for cost-aware active selection when features bear different acquisition costs. The effectiveness of the proposed approach is well validated by both theoretical analysis and experimental study.

Deep convolutional neural networks have achieved great success in various applications. However, training an effective DNN model for a specific task is rather challenging because it requires a prior knowledge or experience to design the network architecture, repeated trial-and-error process to tune the parameters, and a large set of labeled data to train the model. In this paper, we propose to overcome these challenges by actively adapting a pre-trained model to a new task with less labeled examples. Specifically, the pre-trained model is iteratively fine tuned based on the most useful examples. The examples are actively selected based on a novel criterion, which jointly estimates the potential contribution of an instance on optimizing the feature representation as well as improving the classification model for the target task. On one hand, the pre-trained model brings plentiful information from its original task, avoiding redesign of the network architecture or training from scratch; and on the other hand, the labeling cost can be significantly reduced by active label querying. Experiments on multiple datasets and different pre-trained models demonstrate that the proposed approach can achieve cost-effective training of DNNs.

Linear classifier is an essential part of machine learning, and improving its robustness has attracted much effort. Logistic regression (LR) is one of the most widely used linear classifier for its simplicity and probabilistic output. To reduce the risk of overfitting, LR was enhanced by introducing a generalized logistic loss (GLL) with a L2-norm regularization, aiming to maximize the minimum margin. However, the strategy of maximizing minimal margin is less robust to noisy data. In this paper, we incorporate GLL with margin distribution to exploit the statistical information from the training data, and propose a margin distribution logistic machine (MDLM) for better generalization performance and robustness. Furthermore, we extend MDLM to a multi-class version and learn different classes simultaneously by utilizing more information shared across these classes. Extensive experimental results validate the effectiveness of MDLM on both binary classification and multi-class classification.

In multi-label learning, it is rather expensive to label instances since they are simultaneously associated with multiple labels. Therefore, active learning, which reduces the labeling cost by actively querying the labels of the most valuable data, becomes particularly important for multi-label learning. A strong multi-label active learning algorithm usually consists of two crucial elements: a reasonable criterion to evaluate the gain of queried label, and an effective classification model, based on whose prediction the criterion can be accurately computed. In this paper, we first introduce an effective multi-label classification model by combining label ranking with threshold learning, which is incrementally trained to avoid retraining from scratch after every query. Based on this model, we then propose to exploit both uncertainty and diversity in the instance space as well as the label space, and actively query the instance-label pairs which can improve the classification model most. Experimental results demonstrate the superiority of the proposed approach to state-of-the-art methods.

Multi-label learning arises in many real-world tasks where an object is naturally associated with multiple concepts. It is well-accepted that, in order to achieve a good performance, the relationship among labels should be exploited. Most existing approaches require the label relationship as prior knowledge, or exploit by counting the label co-occurrence. In this paper, we propose the MAHR approach, which is able to automatically discover and exploit label relationship. Our basic idea is that, if two labels are related, the hypothesis generated for one label can be helpful for the other label. MAHR implements the idea as a boosting approach with a hypothesis reuse mechanism. In each boosting round, the base learner for a label is generated by not only learning on its own task but also reusing the hypotheses from other labels, and the amount of reuse across labels provides an estimate of the label relationship. Extensive experimental results validate that MAHR is able to achieve superior performance and discover reasonable label relationship. Moreover, we disclose that the label relationship is usually asymmetric.