Search

The precision of unsupervised point cloud registration methods is typically limited by the lack of reliable inlier estimation and self-supervised signal, especially in partially overlapping scenarios. In this article, we propose an effective inlier estimation method for unsupervised point cloud registration by capturing geometric structure consistency between the source point cloud and its corresponding reference point cloud copy. Specifically, to obtain a high-quality reference point cloud copy, a one-nearest neighborhood (1-NN) point cloud is generated by input point cloud, which facilitates matching map construction and allows for integrating dual neighborhood matching scores of 1-NN point cloud and input point cloud to improve matching confidence. Benefiting from the high-quality reference copy, we argue that the neighborhood graph formed by inlier and its neighborhood should have consistency between source point cloud and its corresponding reference copy. Based on this observation, we construct transformation-invariant geometric structure representations and capture geometric structure consistency to score the inlier confidence for estimated correspondences between source point cloud and its reference copy. This strategy can simultaneously provide the reliable self-supervised signals for model optimization. Finally, we further calculate transformation estimation by the weighted SVD algorithm with the estimated correspondences and the corresponding inlier confidence. We train the proposed model in an unsupervised manner, and extensive experiments on synthetic and real-world datasets illustrate the effectiveness of the proposed method.

Three-dimensional point cloud registration is an important field in computer vision. Recently, due to the increasingly complex scenes and incomplete observations, many partial-overlap registration methods based on overlap estimation have been proposed. These methods heavily rely on the extracted overlapping regions with their performances greatly degraded when the overlapping region extraction underperforms. To solve this problem, we propose a partial-to-partial registration network (RORNet) to find reliable overlapping representations from the partially overlapping point clouds and use these representations for registration. The idea is to select a small number of key points called reliable overlapping representations from the estimated overlapping points, reducing the side effect of overlap estimation errors on registration. Although it may filter out some inliers, the inclusion of outliers has a much bigger influence than the omission of inliers on the registration task. The RORNet is composed of overlapping points' estimation module and representations' generation module. Different from the previous methods of direct registration after extraction of overlapping areas, RORNet adds the step of extracting reliable representations before registration, where the proposed similarity matrix downsampling method is used to filter out the points with low similarity and retain reliable representations, and thus reduce the side effects of overlap estimation errors on the registration. Besides, compared with previous similarity-based and score-based overlap estimation methods, we use the dual-branch structure to combine the benefits of both, which is less sensitive to noise. We perform overlap estimation experiments and registration experiments on the ModelNet40 dataset, outdoor large scene dataset KITTI, and natural data Stanford Bunny dataset. The experimental results demonstrate that our method is superior to other partial registration methods. Our code is available at https://github.com/superYuezhang/RORNet.

The pandemic of coronavirus disease 2019 (COVID-19) has led to a global public health crisis, which caused millions of deaths and billions of infections, greatly increasing the pressure on medical resources. With the continuous emergence of viral mutations, developing automated tools for COVID-19 diagnosis is highly desired to assist the clinical diagnosis and reduce the tedious workload of image interpretation. However, medical images in a single site are usually of a limited amount or weakly labeled, while integrating data scattered around different institutions to build effective models is not allowed due to data policy restrictions. In this article, we propose a novel privacy-preserving cross-site framework for COVID-19 diagnosis with multimodal data, seeking to effectively leverage heterogeneous data from multiple parties while preserving patients' privacy. Specifically, a Siamese branched network is introduced as the backbone to capture inherent relationships across heterogeneous samples. The redesigned network is capable of handling semisupervised inputs in multimodalities and conducting task-specific training, in order to improve the model performance of various scenarios. The framework achieves significant improvement compared with state-of-the-art methods, as we demonstrate through extensive simulations on real-world datasets.

Multiparty learning (MPL) is an emerging framework for privacy-preserving collaborative learning. It enables individual devices to build a knowledge-shared model and remaining sensitive data locally. However, with the continuous increase of users, the heterogeneity gap between data and equipment becomes wider, which leads to the problem of model heterogeneous. In this article, we concentrate on two practical issues: data heterogeneous problem and model heterogeneous problem, and propose a novel personal MPL method named device-performance-driven heterogeneous MPL (HMPL). First, facing the data heterogeneous problem, we focus on the problem of various devices holding arbitrary data sizes. We introduce a heterogeneous feature-map integration method to adaptively unify the various feature maps. Meanwhile, to handle the model heterogeneous problem, as it is essential to customize models for adapting to the various computing performances, we propose a layer-wise model generation and aggregation strategy. The method can generate customized models based on the device's performance. In the aggregation process, the shared model parameters are updated through the rules that the network layers with the same semantics are aggregated with each other. Extensive experiments are conducted on four popular datasets, and the result demonstrates that our proposed framework outperforms the state of the art (SOTA).

Hyperedge prediction aims to predict common relations among multiple nodes that will occur in the future or remain undiscovered in the current hypergraph. It is traditionally modeled as a classification task, which performs hypergraph feature learning and classifies the target samples as either present or absent. However, these approaches involve two issues: (i) in hyperedge feature learning, they fail to measure the influence of nodes on the hyperedges that include them and the neighboring hyperedges, and (ii) in the binary classification task, the quality of the generated negative samples directly impacts the prediction results. To this end, we propose a Hypergraph Contrastive Attention Network (HCAN) model for hyperedge prediction. Inspired by the brain organization, HCAN considers the influence of hyperedges with different orders through the order propagation attention mechanism. It also utilizes the contrastive mechanism to measure the reliability of attention effectively. Furthermore, we design a negative sample generator to produce three different types of negative samples. We evaluate the impact of various negative samples on the model and analyze the problems of binary classification modeling. The effectiveness of HCAN in hyperedge prediction is validated by experimentally comparing 12 baselines on 9 datasets. Our implementations will be publicly available at https://github.com/jianruichen/HCAN., Competing Interests: Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper., (Copyright © 2024 Elsevier Ltd. All rights reserved.)

Multiparty learning provides solutions for training joint models with decentralized data under legal and practical constraints. However, traditional multiparty learning approaches are confronted with obstacles, such as system heterogeneity, statistical heterogeneity, and incentive design. Determining how to deal with these challenges and further improve the efficiency and performance of multiparty learning has become an urgent problem to be solved. In this article, we propose a novel contrastive multiparty learning framework for knowledge refinement and sharing with an accountable incentive mechanism. Since the existing parameter averaging method is contradictory to the learning paradigm of neural networks, we simulate the process of human cognition and communication and analogize multiparty learning as a many-to-one knowledge-sharing problem. The approach is capable of integrating the acquired explicit knowledge of each client in a transparent manner without privacy disclosure, and it reduces the dependence on data distribution and communication environments. The proposed scheme achieves significant improvement in model performance in a variety of scenarios, as we demonstrated through experiments on several real-world datasets.

We explore the effect of geometric structure descriptors on extracting reliable correspondences and obtaining accurate registration for point cloud registration. The point cloud registration task involves the estimation of rigid transformation motion in unorganized point cloud, hence it is crucial to capture the contextual features of the geometric structure in point cloud. Recent coordinates-only methods ignore numerous geometric information in the point cloud which weaken ability to express the global context. We propose Enhanced Geometric Structure Transformer to learn enhanced contextual features of the geometric structure in point cloud and model the structure consistency between point clouds for extracting reliable correspondences, which encodes three explicit enhanced geometric structures and provides significant cues for point cloud registration. More importantly, we report empirical results that Enhanced Geometric Structure Transformer can learn meaningful geometric structure features using none of the following: (i) explicit positional embeddings, (ii) additional feature exchange module such as cross-attention, which can simplify network structure compared with plain Transformer. Extensive experiments on the synthetic dataset and real-world datasets illustrate that our method can achieve competitive results.

Nowadays, pruning techniques have drawn attention to convolutional neural networks (CNNs) for reducing the consumption of computation resources. In particular, the Taylor-based method simplifies the evaluation of importance for each filter as the product of the gradient and weight value of the output features, which outperforms other methods in reductions of parameters and floating point operations (FLOPs). However, the Taylor-based method sacrifices too much accuracy when the overall pruning rate is relatively large compared with other pruning algorithms. In this article, we propose a self-adaptive attention factor (SAAF) to improve the performance of the slimmed model when conventional Taylor-based pruning is utilized under higher pruning. Specifically, SAAF can be calculated by leveraging the remaining ratio of filters at the early pruning stage of the Taylor-based method, and then, some pruned filters can be recovered for improving the accuracy of the slimmed model in terms of SAAF. It means that SAAF can protect filters from being overslimmed to eliminate the degeneration of Taylor-based pruning when the pruning rate is large as well as can compress models apparently across various datasets. We test the efficiency of SAAF on VGG-16 and ResNet-50 with CIFAR-10, Tiny-ImageNet, ImageNet-1000, and remote sensing images. Our method outperforms the traditional Taylor-based method obviously in accuracy, and there are only tiny sacrifices in the reduction of parameters and FLOPs, which is better than other pruning methods.

With the rapid development of remote sensing (RS) technology, high-resolution RS image change detection (CD) has been widely used in many applications. Pixel-based CD techniques are maneuverable and widely used, but vulnerable to noise interference. Object-based CD techniques can effectively utilize the abundant spectrum, texture, shape, and spatial information but easy-to-ignore details of RS images. How to combine the advantages of pixel-based methods and object-based methods remains a challenging problem. Besides, although supervised methods have the capability to learn from data, the true labels representing changed information of RS images are often hard to obtain. To address these issues, this article proposes a novel semisupervised CD framework for high-resolution RS images, which employs small amounts of true labeled data and a lot of unlabeled data to train the CD network. A bihierarchical feature aggregation and extraction network (BFAEN) is designed to achieve the pixelwise together with objectwise feature concatenation feature representation for the comprehensive utilization of the two-level features. In order to alleviate the coarseness and insufficiency of labeled samples, a confident learning algorithm is used to eliminate noisy labels and a novel loss function is designed for training the model using true- and pseudo-labels in a semisupervised fashion. Experimental results on real datasets demonstrate the effectiveness and superiority of the proposed method.

The empirical studies of most existing graph neural networks (GNNs) broadly take the original node feature and adjacency relationship as single-channel input, ignoring the rich information of multiple graph channels. To circumvent this issue, the multichannel graph analysis framework has been developed to fuse graph information across channels. How to model and integrate shared (i.e., consistency) and channel-specific (i.e., complementarity) information is a key issue in multichannel graph analysis. In this article, we propose a cross-channel graph information bottleneck (CCGIB) principle to maximize the agreement for common representations and the disagreement for channel-specific representations. Under this principle, we formulate the consistency and complementarity information bottleneck (IB) objectives. To enable optimization, a viable approach involves deriving variational lower bound and variational upper bound (VarUB) of mutual information terms, subsequently focusing on optimizing these variational bounds to find the approximate solutions. However, obtaining the lower bounds of cross-channel mutual information objectives proves challenging through direct utilization of variational approximation, primarily due to the independence of the distributions. To address this challenge, we leverage the inherent property of joint distributions and subsequently derive variational bounds to effectively optimize these information objectives. Extensive experiments on graph benchmark datasets demonstrate the superior effectiveness of the proposed method.

The few-shot image classification task is to enable a model to identify novel classes by using only a few labeled samples as references. In general, the more knowledge a model has, the more robust it is when facing novel situations. Although directly introducing large amounts of new training data to acquire more knowledge is an attractive solution, it violates the purpose of few-shot learning with respect to reducing dependence on big data. Another viable option is to enable the model to accumulate knowledge more effectively from existing data, i.e., improve the utilization of existing data. In this article, we propose a new data augmentation method called self-mixup (SM) to assemble different augmented instances of the same image, which facilitates the model to more effectively accumulate knowledge from limited training data. In addition to the utilization of data, few-shot learning faces another challenge related to feature extraction. Specifically, existing metric-based few-shot classification methods rely on comparing the extracted features of the novel classes, but the widely adopted downsampling structures in various networks can lead to feature degradation due to the violation of the sampling theorem, and the degraded features are not conducive to robust classification. To alleviate this problem, we propose a calibration-adaptive downsampling (CADS) that calibrates and utilizes the characteristics of different features, which can facilitate robust feature extraction and benefit classification. By improving data utilization and feature extraction, our method shows superior performance on four widely adopted few-shot classification datasets.

It is demonstrated that higher-order patterns beyond pairwise relations can significantly enhance the learning capability of existing graph-based models, and simplex is one of the primary form for graphically representing higher-order patterns. Predicting unknown (disappeared) simplices in real-world complex networks can provide us with deeper insights, thereby assisting us in making better decisions. Nevertheless, previous efforts to predict simplices suffer from two issues: (i) they mainly focus on 2- or 3-simplices, and there are few models available for predicting simplices of arbitrary orders, and (ii) they lack the ability to analyze and learn the features of simplices from the perspective of dynamics. In this paper, we present a Higher-order Neurodynamical Equation for Simplex Prediction of arbitrary order (HNESP), which is a framework that combines neural networks and neurodynamics. Specifically, HNESP simulates the dynamical coupling process of nodes in simplicial complexes through different relations (i.e., strong pairwise relation, weak pairwise relation, and simplex) to learn node-level representations, while explaining the learning mechanism of neural networks from neurodynamics. To enrich the higher-order information contained in simplices, we exploit the entropy and normalized multivariate mutual information of different sub-structures of simplices to acquire simplex-level representations. Furthermore, simplex-level representations and multi-layer perceptron are used to quantify the existence probability of simplices. The effectiveness of HNESP is demonstrated by extensive simulations on seven higher-order benchmarks. Experimental results show that HNESP improves the AUC values of the state-of-the-art baselines by an average of 8.32%. Our implementations will be publicly available at: https://github.com/jianruichen/HNESP., Competing Interests: Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper., (Copyright © 2024 Elsevier Ltd. All rights reserved.)

Multiobjective multitasking optimization (MTO) needs to solve a set of multiobjective optimization problems simultaneously, and tries to speed up their solution by transferring useful search experiences across tasks. However, the quality of transfer solutions will significantly impact the transfer effect, which may even deteriorate the optimization performance with an improper selection of transfer solutions. To alleviate this issue, this article suggests a new multiobjective multitasking evolutionary algorithm (MMTEA) with decomposition-based transfer selection, called MMTEA-DTS. In this algorithm, all tasks are first decomposed into a set of subproblems, and then the transfer potential of each solution can be quantified based on the performance improvement ratio of its associated subproblem. Only high-potential solutions are selected to promote knowledge transfer. Moreover, to diversify the transfer of search experiences, a hybrid transfer evolution method is designed in this article. In this way, more diverse search experiences are transferred from high-potential solutions across different tasks to speed up their convergence. Three well-known benchmark suites suggested in the competition of evolutionary MTO and one real-world problem suite are used to verify the effectiveness of MMTEA-DTS. The experiments validate its advantages in solving most of the test problems when compared to five recently proposed MMTEAs.

The multi-layer network consists of the interactions between different layers, where each layer of the network is depicted as a graph, providing a comprehensive way to model the underlying complex systems. The layer-specific modules of multi-layer networks are critical to understanding the structure and function of the system. However, existing methods fail to characterize and balance the connectivity and specificity of layer-specific modules in networks because of the complicated inter- and intra-coupling of various layers. To address the above issues, a joint learning graph clustering algorithm (DRDF) for detecting layer-specific modules in multi-layer networks is proposed, which simultaneously learns the deep representation and discriminative features. Specifically, DRDF learns the deep representation with deep nonnegative matrix factorization, where the high-order topology of the multi-layer network is gradually and precisely characterized. Moreover, it addresses the specificity of modules with discriminative feature learning, where the intra-class compactness and inter-class separation of pseudo-labels of clusters are explored as self-supervised information, thereby providing a more accurate method to explicitly model the specificity of the multi-layer network. Finally, DRDF balances the connectivity and specificity of layer-specific modules with joint learning, where the overall objective of the graph clustering algorithm and optimization rules are derived. The experiments on ten multi-layer networks showed that DRDF not only outperforms eight baselines on graph clustering but also enhances the robustness of algorithms., Competing Interests: Declaration of competing interest We declare that we have no financial and personal relationships with other people or organizations that can inappropriately influence our work, there is no professional or other personal interest of any nature or kind in any product, service and/or company that could be construed as influencing the position presented in, or the review of, the manuscript entitled., (Copyright © 2023 Elsevier Ltd. All rights reserved.)

Graph neural networks (GNNs) have demonstrated great success in many graph data-based applications. The impressive behavior of GNNs typically relies on the availability of a sufficient amount of labeled data for model training. However, in practice, obtaining a large number of annotations is prohibitively labor-intensive and even impossible. Co-training is a popular semi-supervised learning (SSL) paradigm, which trains multiple models based on a common training set while augmenting the limited amount of labeled data used for training each model via the pseudolabeled data generated from the prediction results of other models. Most of the existing co-training works do not control the quality of pseudolabeled data when using them. Therefore, the inaccurate pseudolabels generated by immature models in the early stage of the training process are likely to cause noticeable errors when they are used for augmenting the training data for other models. To address this issue, we propose a self-paced co-training for the GNN (SPC-GNN) framework for semi-supervised node classification. This framework trains multiple GNNs with the same or different structures on different representations of the same training data. Each GNN carries out SSL by using both the originally available labeled data and the augmented pseudolabeled data generated from other GNNs. To control the quality of pseudolabels, a self-paced label augmentation strategy is designed to make the pseudolabels generated at a higher confidence level to be utilized earlier during training such that the negative impact of inaccurate pseudolabels on training data augmentation, and accordingly, the subsequent training process can be mitigated. Finally, each of the trained GNN is evaluated on a validation set, and the best-performing one is chosen as the output. To improve the training effectiveness of the framework, we devise a pretraining followed by a two-step optimization scheme to train GNNs. Experimental results on the node classification task demonstrate that the proposed framework achieves significant improvement over the state-of-the-art SSL methods.

Multi-party learning provides an effective approach for training a machine learning model, e.g., deep neural networks (DNNs), over decentralized data by leveraging multiple decentralized computing devices, subjected to legal and practical constraints. Different parties, so-called local participants, usually provide heterogenous data in a decentralized mode, leading to non-IID data distributions across different local participants which pose a notorious challenge for multi-party learning. To address this challenge, we propose a novel heterogeneous differentiable sampling (HDS) framework. Inspired by the dropout strategy in DNNs, a data-driven network sampling strategy is devised in the HDS framework, with differentiable sampling rates which allow each local participant to extract from a common global model the optimal local model that best adapts to its own data properties so that the size of the local model can be significantly reduced to enable more efficient inference. Meanwhile, co-adaptation of the global model via learning such local models allows for achieving better learning performance under non-IID data distributions and speeds up the convergence of the global model. Experiments have demonstrated the superiority of the proposed method over several popular multi-party learning techniques in the multi-party settings with non-IID data distributions.

Graph classification aims to predict the label associated with a graph and is an important graph analytic task with widespread applications. Recently, graph neural networks (GNNs) have achieved state-of-the-art results on purely supervised graph classification by virtue of the powerful representation ability of neural networks. However, almost all of them ignore the fact that graph classification usually lacks reasonably sufficient labeled data in practical scenarios due to the inherent labeling difficulty caused by the high complexity of graph data. The existing semisupervised GNNs typically focus on the task of node classification and are incapable to deal with graph classification. To tackle the challenging but practically useful scenario, we propose a novel and general semisupervised GNN framework for graph classification, which takes full advantage of a slight amount of labeled graphs and abundant unlabeled graph data. In our framework, we train two GNNs as complementary views for collaboratively learning high-quality classifiers using both labeled and unlabeled graphs. To further exploit the view itself, we constantly select pseudo-labeled graph examples with high confidence from its own view for enlarging the labeled graph dataset and enhancing predictions on graphs. Furthermore, the proposed framework is investigated on two specific implementation regimes with a few labeled graphs and the extremely few labeled graphs, respectively. Extensive experimental results demonstrate the effectiveness of our proposed semisupervised GNN framework for graph classification on several benchmark datasets.

It is attractive to extract plausible 3-D information from a single 2-D image, and self-supervised learning has shown impressive potential in this field. However, when only monocular videos are available as training data, moving objects at similar speeds to the camera can disturb the reprojection process during training. Existing methods filter out some moving pixels by comparing pixelwise photometric error, but the illumination inconsistency between frames leads to incomplete filtering. In addition, existing methods calculate photometric error within local windows, which leads to the fact that even if an anomalous pixel is masked out, it can still implicitly disturb the reprojection process, as long as it is in the local neighborhood of a nonanomalous pixel. Moreover, the ill-posed nature of monocular depth estimation makes the same scene correspond to multiple plausible depth maps, which damages the robustness of the model. In order to alleviate the above problems, we propose: 1) a self-reprojection mask to further filter out moving objects while avoiding illumination inconsistency; 2) a self-statistical mask method to prevent the filtered anomalous pixels from implicitly disturbing the reprojection; and 3) a self-distillation augmentation consistency loss to reduce the impact of ill-posed nature of monocular depth estimation. Our method shows superior performance on the KITTI dataset, especially when evaluating only the depth of potential moving objects.

Audio and vision are two main modalities in video data. Multimodal learning, especially for audiovisual learning, has drawn considerable attention recently, which can boost the performance of various computer vision tasks. However, in video summarization, most existing approaches just exploit the visual information while neglecting the audio information. In this brief, we argue that the audio modality can assist vision modality to better understand the video content and structure and further benefit the summarization process. Motivated by this, we propose to jointly exploit the audio and visual information for the video summarization task and develop an audiovisual recurrent network (AVRN) to achieve this. Specifically, the proposed AVRN can be separated into three parts: 1) the two-stream long-short term memory (LSTM) is used to encode the audio and visual feature sequentially by capturing their temporal dependency; 2) the audiovisual fusion LSTM is used to fuse the two modalities by exploring the latent consistency between them; and 3) the self-attention video encoder is adopted to capture the global dependency in the video. Finally, the fused audiovisual information and the integrated temporal and global dependencies are jointly used to predict the video summary. Practically, the experimental results on the two benchmarks, i.e., SumMe and TVsum, have demonstrated the effectiveness of each part and the superiority of AVRN compared with those approaches just exploiting visual information for video summarization.

Complex systems in nature and society consist of various types of interactions, where each type of interaction belongs to a layer, resulting in the so-called multilayer networks. Identifying specific modules for each layer is of great significance for revealing the structure-function relations in multilayer networks. However, the available approaches are criticized undesirable because they fail to explicitly the specificity of modules, and balance the specificity and connectivity of modules. To overcome these drawbacks, we propose an accurate and flexible algorithm by joint learning matrix factorization and sparse representation (jMFSR) for specific modules in multilayer networks, where matrix factorization extracts features of vertices and sparse representation discovers specific modules. To exploit the discriminative latent features of vertices in multilayer networks, jMFSR incorporates linear discriminant analysis (LDA) into non-negative matrix factorization (NMF) to learn features of vertices that distinguish the categories. To explicitly measure the specificity of features, jMFSR decomposes features of vertices into common and specific parts, thereby enhancing the quality of features. Then, jMFSR jointly learns feature extraction, common-specific feature factorization, and clustering of multilayer networks. The experiments on 11 datasets indicate that jMFSR significantly outperforms state-of-the-art baselines in terms of various measurements.

The performance of machine learning algorithms heavily relies on the availability of a large amount of training data. However, in reality, data usually reside in distributed parties such as different institutions and may not be directly gathered and integrated due to various data policy constraints. As a result, some parties may suffer from insufficient data available for training machine learning models. In this article, we propose a multiparty dual learning (MPDL) framework to alleviate the problem of limited data with poor quality in an isolated party. Since the knowledge-sharing processes for multiple parties always emerge in dual forms, we show that dual learning is naturally suitable to handle the challenge of missing data, and explicitly exploits the probabilistic correlation and structural relationship between dual tasks to regularize the training process. We introduce a feature-oriented differential privacy with mathematical proof, in order to avoid possible privacy leakage of raw features in the dual inference process. The approach requires minimal modifications to the existing multiparty learning structure, and each party can build flexible and powerful models separately, whose accuracy is no less than nondistributed self-learning approaches. The MPDL framework achieves significant improvement compared with state-of-the-art multiparty learning methods, as we demonstrated through simulations on real-world datasets.

An expensive multimodal optimization problem (EMMOP) is that the computation of the objective function is time consuming and it has multiple global optima. This article proposes a decomposition differential evolution (DE) based on the radial basis function (RBF) for EMMOPs, called D/REM. It mainly consists of two phases: the promising subregions detection (PSD) and the local search phase (LSP). In PSD, a population update strategy is designed and the mean-shift clustering is employed to predict the promising subregions of EMMOP. In LSP, a local RBF surrogate model is constructed for each promising subregion and each local RBF surrogate model tracks a global optimum of EMMOP. In this way, an EMMOP is decomposed into many expensive global optimization subproblems. To handle these subproblems, a popular DE variant, JADE, acts as the search engine to deal with these subproblems. A large number of numerical experiments unambiguously validate that D/REM can solve EMMOPs effectively and efficiently.

Temporal networks are ubiquitous in nature and society, and tracking the dynamics of networks is fundamental for investigating the mechanisms of systems. Dynamic communities in temporal networks simultaneously reflect the topology of the current snapshot (clustering accuracy) and historical ones (clustering drift). Current algorithms are criticized for their inability to characterize the dynamics of networks at the vertex level, independence of feature extraction and clustering, and high time complexity. In this study, we solve these problems by proposing a novel joint learning model for dynamic community detection in temporal networks (also known as jLMDC) via joining feature extraction and clustering. This model is formulated as a constrained optimization problem. Vertices are classified into dynamic and static groups by exploring the topological structure of temporal networks to fully exploit their dynamics at each time step. Then, jLMDC updates the features of dynamic vertices by preserving features of static ones during optimization. The advantage of jLMDC is that features are extracted under the guidance of clustering, promoting performance, and saving the running time of the algorithm. Finally, we extend jLMDC to detect the overlapping dynamic community in temporal networks. The experimental results on 11 temporal networks demonstrate that jLMDC improves accuracy up to 8.23% and saves 24.89% of running time on average compared to state-of-the-art methods.

Non-negative matrix factorization (NMF) is one of the most popular techniques for data representation and clustering and has been widely used in machine learning and data analysis. NMF concentrates the features of each sample into a vector and approximates it by the linear combination of basis vectors, such that the low-dimensional representations are achieved. However, in real-world applications, the features usually have different importance. To exploit the discriminative features, some methods project the samples into the subspace with a transformation matrix, which disturbs the original feature attributes and neglects the diversity of samples. To alleviate the above problems, we propose the feature weighted NMF (FNMF) in this article. The salient properties of FNMF can be summarized as three-fold: 1) it learns the weights of features adaptively according to their importance; 2) it utilizes multiple feature weighting components to preserve the diversity; and 3) it can be solved efficiently with the suggested optimization algorithm. The performance on synthetic and real-world datasets demonstrates that the proposed method obtains the state-of-the-art performance.

Cardinality constraint, namely, constraining the number of nonzero outputs of models, has been widely used in structural learning. It can be used for modeling the dependencies between multidimensional labels. In hashing, the final outputs are also binary codes, which are similar to multidimensional labels. It has been validated that estimating how many 1's in a multidimensional label vector is easier than directly predicting which elements are 1 and estimating cardinality as a prior step will improve the classification performance. Hence, in this article, we incorporate cardinality constraint into the unsupervised image hashing problem. The proposed model is divided into two steps: 1) estimating the cardinalities of hashing codes and 2) then estimating which bits are 1. Unlike multidimensional labels that are known and fixed in the training phase, the hashing codes are generally learned through an iterative method and, therefore, their cardinalities are unknown and not fixed during the learning procedure. We use a neural network as a cardinality predictor and its parameters are jointly learned with the hashing code generator, which is an autoencoder in our model. The experiments demonstrate the efficiency of our proposed method.

Recently, stochastic hard thresholding (HT) optimization methods [e.g., stochastic variance reduced gradient hard thresholding (SVRGHT)] are becoming more attractive for solving large-scale sparsity/rank-constrained problems. However, they have much higher HT oracle complexities, especially for high-dimensional data or large-scale matrices. To address this issue and inspired by the well-known Gradient Support Pursuit (GraSP) method, this article proposes a new Relaxed Gradient Support Pursuit (RGraSP) framework. Unlike GraSP, RGraSP only requires to yield an approximation solution at each iteration. Based on the property of RGraSP, we also present an efficient stochastic variance reduction-gradient support pursuit algorithm and its fast version (called stochastic variance reduced gradient support pursuit (SVRGSP+). We prove that the gradient oracle complexity of both our algorithms is two times less than that of SVRGHT. In particular, their HT complexity is about κ ∧s times less than that of SVRGHT, where κ ∧s is the restricted condition number. Moreover, we prove that our algorithms enjoy fast linear convergence to an approximately global optimum, and also present an asynchronous parallel variant to deal with very high-dimensional and sparse data. Experimental results on both synthetic and real-world datasets show that our algorithms yield superior results than the state-of-the-art gradient HT methods.

As a unified framework for graph neural networks, message passing-based neural network (MPNN) has attracted a lot of research interest and has been shown successfully in a number of domains in recent years. However, because of over-smoothing and vanishing gradients, deep MPNNs are still difficult to train. To alleviate these issues, we first introduce a deep hierarchical layer aggregation (DHLA) strategy, which utilizes a block-based layer aggregation to aggregate representations from different layers and transfers the output of the previous block to the subsequent block, so that deeper MPNNs can be easily trained. Additionally, to stabilize the training process, we also develop a novel normalization strategy, neighbor normalization (NeighborNorm), which normalizes the neighbor of each node to further address the training issue in deep MPNNs. Our analysis reveals that NeighborNorm can smooth the gradient of the loss function, i.e., adding NeighborNorm makes the optimization landscape much easier to navigate. Experimental results on two typical graph pattern-recognition tasks, including node classification and graph classification, demonstrate the necessity and effectiveness of the proposed strategies for graph message-passing neural networks.

Network representation learning or embedding aims to project the network into a low-dimensional space that can be devoted to different network tasks. Temporal networks are an important type of network whose topological structure changes over time. Compared with methods on static networks, temporal network embedding (TNE) methods are facing three challenges: 1) it cannot describe the temporal dependence across network snapshots; 2) the node embedding in the latent space fails to indicate changes in the network topology; and 3) it cannot avoid a lot of redundant computation via parameter inheritance on a series of snapshots. To overcome these problems, we propose a novel TNE method named temporal network embedding method based on the VAE framework (TVAE), which is based on a variational autoencoder (VAE) to capture the evolution of temporal networks for link prediction. It not only generates low-dimensional embedding vectors for nodes but also preserves the dynamic nonlinear features of temporal networks. Through the combination of a self-attention mechanism and recurrent neural networks, TVAE can update node representations and keep the temporal dependence of vectors over time. We utilize parameter inheritance to keep the new embedding close to the previous one, rather than explicitly using regularization, and thus, it is effective for large-scale networks. We evaluate our model and several baselines on synthetic data sets and real-world networks. The experimental results demonstrate that TVAE has superior performance and lower time cost compared with the baselines.

The searching ability of the population-based search algorithms strongly relies on the coordinate system on which they are implemented. However, the widely used coordinate systems in the existing multifactorial optimization (MFO) algorithms are still fixed and might not be suitable for various function landscapes with differential modalities, rotations, and dimensions; thus, the intertask knowledge transfer might not be efficient. Therefore, this article proposes a novel intertask knowledge transfer strategy for MFOs implemented upon an active coordinate system that is established on a common subspace of two search spaces. The proper coordinate system might identify some common modality in a proper subspace to some extent. In this article, to seek the intermediate subspace, we innovatively introduce the geodesic flow that starts from a subspace, reaching another subspace in unit time. A low-dimension intermediate subspace is drawn from a uniform distribution defined on the geodesic flow, and the corresponding coordinate system is given. The intertask trial generation method is applied to the individuals by first projecting them on the low-dimension subspace, which reveals the important invariant features of the multiple function landscapes. Since intermediate subspace is generated from the major eigenvectors of tasks' spaces, this model turns out to be intrinsically regularized by neglecting the minor and small eigenvalues. Therefore, the transfer strategy can alleviate the influence of noise led by redundant dimensions. The proposed method exhibits promising performance in the experiments.

With the presence of complex background noise, parasitic light, and dust attachment, it is still a challenging issue to perform high-precision laser-induced damage change detection of optical elements in the captured optical images. For resolving this problem, this paper presents an end-to-end damage change detection model based on siamese network and multi-layer perceptrons (SiamMLP). Firstly, representative features of bi-temporal damage images are efficiently extracted by the cascaded multi-layer perceptron modules in the siamese network. After that, the extracted features are concatenated and then classified into changed and unchanged classes. Due to its concise architecture and strong feature representation ability, the proposed method obtains excellent damage change detection results efficiently and effectively. To address the unbalanced distribution of hard and easy samples, a novel metric called hard metric is introduced in this paper for quantitatively evaluating the classification difficulty degree of the samples. The hard metric assigns a classification difficulty for each individual sample to precisely adjust the loss assigned to the sample. In the training stage, a novel hard loss is presented to train the proposed model. Cooperating with the hard metric, the hard loss can up-weight the loss of hard samples and down-weight the loss of easy samples, which results in a more powerful online hard sample mining ability of the proposed model. The experimental results on two real datasets validate the effectiveness and superiority of the proposed method.

Performance of multi/many-objective evolutionary algorithms (MOEAs) based on decomposition is highly impacted by the Pareto front (PF) shapes of multi/many-objective optimization problems (MOPs), as their adopted weight vectors may not properly fit the PF shapes. To avoid this mismatch, some MOEAs treat solutions as weight vectors to guide the evolutionary search, which can adapt to the target MOP's PF automatically. However, their performance is still affected by the similarity metric used to select weight vectors. To address this issue, this article proposes a fuzzy decomposition-based MOEA. First, a fuzzy prediction is designed to estimate the population's shape, which helps to exactly reflect the similarities of solutions. Then, N least similar solutions are extracted as weight vectors to obtain N constrained fuzzy subproblems ( N is the population size), and accordingly, a shared weight vector is calculated for all subproblems to provide a stable search direction. Finally, the corner solution for each of m least similar subproblems ( m is the objective number) is preserved to maintain diversity, while one solution having the best aggregated value on the shared weight vector is selected for each of the remaining subproblems to speed up convergence. When compared to several competitive MOEAs in solving a variety of test MOPs, the proposed algorithm shows some advantages at fitting their different PF shapes.

Recently, deep-learning-based feature extraction (FE) methods have shown great potential in hyperspectral image (HSI) processing. Unfortunately, it also brings a challenge that the training of the deep learning networks always requires large amounts of labeled samples, which is hardly available for HSI data. To address this issue, in this article, a novel unsupervised deep-learning-based FE method is proposed, which is trained in an end-to-end style. The proposed framework consists of an encoder subnetwork and a decoder subnetwork. The structure of the two subnetworks is symmetric for obtaining better downsampling and upsampling representation. Considering both spectral and spatial information, 3-D all convolution nets and deconvolution nets are used to structure the encoder subnetwork and decoder subnetwork, respectively. However, 3-D convolution and deconvolution kernels bring more parameters, which can deteriorate the quality of the obtained features. To alleviate this problem, a novel cost function with a sparse regular term is designed to obtain more robust feature representation. Experimental results on publicly available datasets indicate that the proposed method can obtain robust and effective features for subsequent classification tasks.

Artificial neural networks inspired from the learning mechanism of the brain have achieved great successes in machine learning, especially those with deep layers. The commonly used neural networks follow the hierarchical multilayer architecture with no connections between nodes in the same layer. In this article, we propose a new group architectures for neural-network learning. In the new architecture, the neurons are assigned irregularly in a group and a neuron may connect to any neurons in the group. The connections are assigned automatically by optimizing a novel connecting structure learning probabilistic model which is established based on the principle that more relevant input and output nodes deserve a denser connection between them. In order to efficiently evolve the connections, we propose to directly model the architecture without involving weights and biases which significantly reduce the computational complexity of the objective function. The model is optimized via an improved particle swarm optimization algorithm. After the architecture is optimized, the connecting weights and biases are then determined and we find the architecture is robust to corruptions. From experiments, the proposed architecture significantly outperforms existing popular architectures on noise-corrupted images when trained only by pure images.

In the past decade, deep learning (DL) has achieved unprecedented success in numerous fields, such as computer vision and healthcare. Particularly, DL is experiencing an increasing development in advanced medical image analysis applications in terms of segmentation, classification, detection, and other tasks. On the one hand, tremendous needs that leverage DL's power for medical image analysis arise from the research community of a medical, clinical, and informatics background to share their knowledge, skills, and experience jointly. On the other hand, barriers between disciplines are on the road for them, often hampering a full and efficient collaboration. To this end, we propose our novel open-source platform, i.e., MEDAS-the MEDical open-source platform As Service. To the best of our knowledge, MEDAS is the first open-source platform providing collaborative and interactive services for researchers from a medical background using DL-related toolkits easily and for scientists or engineers from informatics modeling faster. Based on tools and utilities from the idea of RINV (Rapid Implementation aNd Verification), our proposed platform implements tools in pre-processing, post-processing, augmentation, visualization, and other phases needed in medical image analysis. Five tasks, concerning lung, liver, brain, chest, and pathology, are validated and demonstrated to be efficiently realizable by using MEDAS. MEDAS is available at http://medas.bnc.org.cn/., Competing Interests: Conflict of interestThe authors declare that they have no conflict of interest., (© The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2021.)

Extracting 3D information from a single optical image is very attractive. Recently emerging self-supervised methods can learn depth representations without using ground truth depth maps as training data by transforming the depth prediction task into an image synthesis task. However, existing methods rely on a differentiable bilinear sampler for image synthesis, which results in each pixel in a synthetic image being derived from only four pixels in the source image and causes each pixel in the depth map to perceive only a few pixels in the source image. In addition, when calculating the photometric error between a synthetic image and its corresponding target image, existing methods only consider the photometric error within a small neighborhood of each single pixel and therefore ignore correlations between larger areas, which causes the model to tend to fall into the local optima for small patches. In order to extend the perceptual area of the depth map over the source image, we propose a novel multi-scale method that downsamples the predicted depth map and performs image synthesis at different resolutions, which enables each pixel in the depth map to perceive more pixels in the source image and improves the performance of the model. As for the locality of photometric error, we propose a structural similarity (SSIM) pyramid loss to allow the model to sense the difference between images in multiple areas of different sizes. Experimental results show that our method achieves superior performance on both outdoor and indoor benchmarks.

Face is one of the most attractive sensitive information in visual shared data. It is an urgent task to design an effective face deidentification method to achieve a balance between facial privacy protection and data utilities when sharing data. Most of the previous methods for face deidentification rely on attribute supervision to preserve a certain kind of identity-independent utility but lose the other identity-independent data utilities. In this article, we mainly propose a novel disentangled representation learning architecture for multiple attributes preserving face deidentification called replacing and restoring variational autoencoders (R 2 VAEs). The R 2 VAEs disentangle the identity-related factors and the identity-independent factors so that the identity-related information can be obfuscated, while they do not change the identity-independent attribute information. Moreover, to improve the details of the facial region and make the deidentified face blends into the image scene seamlessly, the image inpainting network is employed to fill in the original facial region by using the deidentified face as a priori. Experimental results demonstrate that the proposed method effectively deidentifies face while maximizing the preservation of the identity-independent information, which ensures the semantic integrity and visual quality of shared images.

Cancer progression is dynamic, and tracking dynamic modules is promising for cancer diagnosis and therapy. Accumulated genomic data provide us an opportunity to investigate the underlying mechanisms of cancers. However, as far as we know, no algorithm has been designed for dynamic modules by integrating heterogeneous omics data. To address this issue, we propose an integrative framework for dynamic module detection based on regularized nonnegative matrix factorization method (DrNMF) by integrating the gene expression and protein interaction network. To remove the heterogeneity of genomic data, we divide the samples of expression profiles into groups to construct gene co-expression networks. To characterize the dynamics of modules, the temporal smoothness framework is adopted, in which the gene co-expression network at the previous stage and protein interaction network are incorporated into the objective function of DrNMF via regularization. The experimental results demonstrate that DrNMF is superior to state-of-the-art methods in terms of accuracy. For breast cancer data, the obtained dynamic modules are more enriched by the known pathways, and can be used to predict the stages of cancers and survival time of patients. The proposed model and algorithm provide an effective integrative analysis of heterogeneous genomic data for cancer progression.

The global alignment of biological networks (GABN) aims to find an optimal alignment between proteins across species, such that both the biological structures and the topological structures of the proteins are maximally conserved. The research on GABN has attracted great attention due to its applications on species evolution, orthology detection and genetic analyses. Most of the existing methods for GABN are difficult to obtain a good tradeoff between the conservation of the biological structures and topological structures. In this paper, we propose a multi-neighborhood learning method for solving GABN (called as CLMNA). CLMNA first models GABN as an optimization of a weighted similarity which evaluates the conserved biological and topological similarities of an alignment, and then it combines a first-proximity, second-proximity and individual-aware proximity learning algorithm to solve the modeled problem. Finally, systematic experiments on 10 pairs of biological networks across 5 species show the superiority of CLMNA over the state-of-the-art network alignment algorithms. They also validate the effectiveness of CLMNA as a refinement method on improving the performance of the compared algorithms.

Graph convolutional networks (GCNs) have been widely used for representation learning on graph data, which can capture structural patterns on a graph via specifically designed convolution and readout operations. In many graph classification applications, GCN-based approaches have outperformed traditional methods. However, most of the existing GCNs are inefficient to preserve local information of graphs - a limitation that is especially problematic for graph classification. In this work, we propose a locality-preserving dense GCN with graph context-aware node representations. Specifically, our proposed model incorporates a local node feature reconstruction module to preserve initial node features into node representations, which is realized via a simple but effective encoder-decoder mechanism. To capture local structural patterns in neighborhoods representing different ranges of locality, dense connectivity is introduced to connect each convolutional layer and its corresponding readout with all previous convolutional layers. To enhance node representativeness, the output of each convolutional layer is concatenated with the output of the previous layer's readout to form a global context-aware node representation. In addition, a self-attention module is introduced to aggregate layer-wise representations to form the final graph-level representation. Experiments on benchmark datasets demonstrate the superiority of the proposed model over state-of-the-art methods in terms of classification accuracy., Competing Interests: Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper., (Copyright © 2021 Elsevier Ltd. All rights reserved.)

Background: Gene prioritization (gene ranking) aims to obtain the centrality of genes, which is critical for cancer diagnosis and therapy since keys genes correspond to the biomarkers or targets of drugs. Great efforts have been devoted to the gene ranking problem by exploring the similarity between candidate and known disease-causing genes. However, when the number of disease-causing genes is limited, they are not applicable largely due to the low accuracy. Actually, the number of disease-causing genes for cancers, particularly for these rare cancers, are really limited. Therefore, there is a critical needed to design effective and efficient algorithms for gene ranking with limited prior disease-causing genes., Results: In this study, we propose a transfer learning based algorithm for gene prioritization (called TLGP) in the cancer (target domain) without disease-causing genes by transferring knowledge from other cancers (source domain). The underlying assumption is that knowledge shared by similar cancers improves the accuracy of gene prioritization. Specifically, TLGP first quantifies the similarity between the target and source domain by calculating the affinity matrix for genes. Then, TLGP automatically learns a fusion network for the target cancer by fusing affinity matrix, pathogenic genes and genomic data of source cancers. Finally, genes in the target cancer are prioritized. The experimental results indicate that the learnt fusion network is more reliable than gene co-expression network, implying that transferring knowledge from other cancers improves the accuracy of network construction. Moreover, TLGP outperforms state-of-the-art approaches in terms of accuracy, improving at least 5%., Conclusion: The proposed model and method provide an effective and efficient strategy for gene ranking by integrating genomic data from various cancers., (© 2021. The Author(s).)

Inspired by the learning rules of humans/animals, self-paced learning (SPL) ranks the samples from easy to complex and assigns real-valued weights to the samples. The current SPL regimes adopt an increasing pace parameter to select the training samples. Obviously, it is difficult to tune the pace parameter during iterations. Furthermore, the current SPL regimes cannot go back to the previous stage even when the current model is worse than the previous one. In this article, a novel Pareto SPL (PSPL) approach is proposed to address the above issues. In PSPL, the SPL problem is considered as a multiobjective optimization problem. Then, a Pareto-based differential evolution algorithm is utilized to optimize the self-paced function and the weighted loss function simultaneously. In PSPL, a new representation method is designed to assign nonpositive weights to the unselected samples. Then, a modified mutation operator inspired by the SPL idea is proposed to generate new population based on ranking the individuals and assigning weights. No pace parameters are introduced in the proposed technique and PSPL can obtain entire solution spectrum to automatically adjust the solution path. The effectiveness of the PSPL has been demonstrated through rigorous studies with matrix factorization and multiclass classification problems.

Retrieving nearest neighbors across correlated data in multiple modalities, such as image-text pairs on Facebook and video-tag pairs on YouTube, has become a challenging task due to the huge amount of data. Multimodal hashing methods that embed data into binary codes can boost the retrieving speed and reduce storage requirements. Since unsupervised multimodal hashing methods are usually inferior to supervised ones and the supervised ones require too much manually labeled data, the proposed method in this paper utilizes partially available labels to design a semisupervised multimodal hashing method. The labels for unlabeled data are treated as an interval type-2 fuzzy set and estimated by the available labels. By defuzzifying the estimated labels using hard partitioning, a supervised multimodal hashing method is used to generate binary codes. Experiments show that the proposed semisupervised method with 50% labels can get a medium performance among the compared supervised ones and achieve approximate performance to the best supervised method with 90% labels. With only 10% labels, the proposed method can still compete with the worst compared supervised one. Furthermore, the proposed label estimation method has been experimentally proven to be more feasible for a multilabeled MIRFlickr data set in a hash lookup task.

Deep neural networks (DNNs), characterized by sophisticated architectures capable of learning a hierarchy of feature representations, have achieved remarkable successes in various applications. Learning DNN's parameters is a crucial but challenging task that is commonly resolved by using gradient-based backpropagation (BP) methods. However, BP-based methods suffer from severe initialization sensitivity and proneness to getting trapped into inferior local optima. To address these issues, we propose a DNN learning framework that hybridizes CC-based optimization with BP-based gradient descent, called BPCC, and implement it by devising a computationally efficient CC-based optimization technique dedicated to DNN parameter learning. In BPCC, BP will intermittently execute for multiple training epochs. Whenever the execution of BP in a training epoch cannot sufficiently decrease the training objective function value, CC will kick in to execute by using the parameter values derived by BP as the starting point. The best parameter values obtained by CC will act as the starting point of BP in its next training epoch. In CC-based optimization, the overall parameter learning task is decomposed into many subtasks of learning a small portion of parameters. These subtasks are individually addressed in a cooperative manner. In this article, we treat neurons as basic decomposition units. Furthermore, to reduce the computational cost, we devise a maturity-based subtask selection strategy to selectively solve some subtasks of higher priority. Experimental results demonstrate the superiority of the proposed method over common-practice DNN parameter learning techniques.

Multi-view graph embedding is aimed at learning low-dimensional representations of nodes that capture various relationships in a multi-view network, where each view represents a type of relationship among nodes. Multitudes of existing graph embedding approaches concentrate on single-view networks, that can only characterize one simple type of proximity relationships among objects. However, most of the real-world complex systems possess multiple types of relationships among entities. In this paper, a novel approach of graph embedding for multi-view networks is proposed, named Multi-view Graph Attention Networks (MGAT). We explore an attention-based architecture for learning node representations from each single view, the network parameters of which are constrained by a novel regularization term. In order to collaboratively integrate multiple types of relationships in different views, a view-focused attention method is explored to aggregate the view-wise node representations. We evaluate the proposed algorithm on several real-world datasets, and it demonstrates that the proposed approach outperforms existing state-of-the-art baselines., Competing Interests: Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper., (Copyright © 2020 Elsevier Ltd. All rights reserved.)

The robust controllability (RC) of a complex system tries to select a set of dominating entities for the functional control of this entire system without uncertain disturbances, and the research on RC will help to understand the system's underlying functions. In this article, we introduce the control cost in signed networks and present a cost-aware robust control (CRC) problem in this scenario. The aim of CRC is to minimize the cost to control a set of dominating nodes and transform a set of unbalanced links into balanced ones, such that the signed network can be robustly controlled without uncertain unbalanced factors (like nodes and links). To solve this problem, we first model CRC as a constrained combination optimization problem, and then present a memetic algorithm with some problem-specific knowledge (like the neighbors of nodes, the constraints of CRC, and the fast computation of the cost under each optimization) to solve this problem on signed networks. The extensive experiments on both real social and biological networks assess that our algorithm outperforms several state-of-the-art RC algorithms.

In this article, a new deep neural network based on sparse filtering and manifold regularization (DSMR) is proposed for feature extraction and classification of polarimetric synthetic aperture radar (PolSAR) data. DSMR uses a novel deep neural network (DNN) to automatically learn features from raw SAR data. During preprocessing, the spatial information between pixels on PolSAR images is exploited to weight each data sample. Then, in the pretraining and fine-tuning, DSMR uses the population sparsity and the lifetime sparsity (dual sparsity) to learn the global features and preserves the local structure of data by neighborhood-based manifold regularization. The dual sparsity only needs to tune a few parameters, and the manifold regularization cuts down the number of training samples. Experimental results on synthesized and real PolSAR data sets from different SAR systems show that DSMR can improve classification accuracy compared with conventional DNNs, even for data sets with a large angle of incidence.

In recent years, deep learning achieves remarkable results in the field of artificial intelligence. However, the training process of deep neural networks may cause the leakage of individual privacy. Given the model and some background information of the target individual, the adversary can maliciously infer the sensitive feature of the target individual. Therefore, it is imperative to preserve the sensitive information in the training data. Differential privacy is a state-of-the-art paradigm for providing the privacy guarantee of datasets, which protects the private and sensitive information from the attack of adversaries significantly. However, the existing privacy-preserving models based on differential privacy are less than satisfactory since traditional approaches always inject the same amount of noise into parameters to preserve the sensitive information, which may impact the trade-off between the model utility and the privacy guarantee of training data. In this paper, we present a general differentially private deep neural networks learning framework based on relevance analysis, which aims to bridge the gap between private and non-private models while providing an effective privacy guarantee of sensitive information. The proposed model perturbs gradients according to the relevance between neurons in different layers and the model output. Specifically, during the process of backward propagation, more noise is added to gradients of neurons that have less relevance to the model output, and vice-versa. Experiments on five real datasets demonstrate that our mechanism not only bridges the gap between private and non-private models, but also prevents the disclosure of sensitive information effectively., (Copyright © 2020 Elsevier Ltd. All rights reserved.)

Image change detection detects the regions of change in multiple images of the same scene taken at different times, which plays a crucial role in many applications. The two most popular image change detection techniques are as follows: pixel-based methods heavily rely on accurate image coregistration while object-based approaches can tolerate coregistration errors to some extent but are sensitive to image segmentation or classification errors. To address these issues, we propose an unsupervised image change detection approach based on a novel bipartite differential neural network (BDNN). The BDNN is a deep neural network with two input ends, which can extract the holistic features from the unchanged regions in the two input images, where two learnable change disguise maps (CDMs) are used to disguise the changed regions in the two input images, respectively, and thus demarcate the unchanged regions therein. The network parameters and CDMs will be learned by optimizing an objective function, which combines a loss function defined as the likelihood of the given input image pair over all possible input image pairs and two constraints imposed on CDMs. Compared with the pixel-based and object-based techniques, the BDNN is less sensitive to inaccurate image coregistration and does not involve image segmentation or classification. In fact, it can even skip over coregistration if the degree of transformation (due to the different view angles and/or positions of the camera) between the two input images is not that large. We compare the proposed approach with several state-of-the-art image change detection methods on various homogeneous and heterogeneous image pairs with and without coregistration. The results demonstrate the superiority of the proposed approach.

With the growing demand for an intelligent system to prevent abnormal events, many methods have been proposed to detect and locate anomalous behaviors in surveillance videos. However, most of these methods contain two shortcomings mainly: distraction of the network and insufficient discriminating ability. In this paper, we propose a local distinguishability aggrandizing network (LDA-Net) in a supervised manner, consisting of a human detection module and an anomaly detection module. In the human detection module, we obtain segmented patches of specific human subjects and take them as the input of the latter module to focus the network on learning motion characteristics of each person. In addition, considering that the auxiliary information, such as the specific type of an action, can aggrandize the whole network to extract distinguishable detail features of normal and abnormal behaviors, the proposed anomaly detection module comprises a primary binary classification sub-branch and an auxiliary distinguishability aggrandizing sub-branch, through which we can jointly detect anomalies and recognize actions. To further reduce the misclassification of the extremely imbalanced datasets, we design a novel inhibition loss function and embed it into the auxiliary sub-branch of the anomaly detection module. Experiments on several public benchmark datasets for frame-level and pixel-level anomaly detection show that the proposed supervised LDA-Net achieves state-of-the-art results on UCSD Ped2 and Subway Exit datasets., (Copyright © 2019 Elsevier Ltd. All rights reserved.)

Networks, such as social networks, biochemical networks, and protein-protein interaction networks are ubiquitous in the real world. Network representation learning aims to embed nodes in a network as low-dimensional, dense, real-valued vectors, and facilitate downstream network analysis. The existing embedding methods commonly endeavor to capture structure information in a network, but lack of consideration of subsequent tasks and synergies between these tasks, which are of equal importance for learning desirable network representations. To address this issue, we propose a novel multi-task network representation learning (MTNRL) framework, which is end-to-end and more effective for underlying tasks. The original network and the incomplete network share a unified embedding layer followed by node classification and link prediction tasks that simultaneously perform on the embedding vectors. By optimizing the multi-task loss function, our framework jointly learns task-oriented embedding representations for each node. Besides, our framework is suitable for all network embedding methods, and the experiment results on several benchmark datasets demonstrate the effectiveness of the proposed framework compared with state-of-the-art methods., (Copyright © 2020 Xie, Jin, Gong, Zhang and Yu.)