Search

How to represent and classify a testing sample for the representation-based classification (RBC) plays an important role in the filed of pattern recognition. As a typical kind of the representation-based classification with promising performance, collaborative representation-based classification (CRC) adopts all the training samples to collaboratively represent and then classify each testing sample with the reconstructive residuals among all the classes. However, most of the CRC methods fail to make full use of the localities and discrimination information of data in collaborative representation. To address this issue to further improve the classification performance, we design a novel supervised CRC method entitled locality-constrained weighted collaborative-competitive representation-based classification (LWCCRC). In the proposed method, the localities of data are taken into account by using the positive and negative nearest samples of each testing sample with their corresponding weighted constraints. Such devised locality-constrained weighted term can model the similarity and natural discrimination information contained in the neighborhood region for each testing sample to obtain the favorable representation. Moreover, a competitive constraint is introduced to enhance pattern discrimination among the categorical collaborative representations. To explore the effectiveness of our proposed LWCCRC, the extensive experiments are carried out on three different types of data sets. The experimental results demonstrate that the proposed LWCCRC significantly outperforms the recent state-of-the-art CRC methods.

The strong feature representation ability of deep learning enables content-based image retrieval (CBIR) to achieve higher retrieval accuracy, while there are still some challenges for CBIR such as high requirements of training labels and retrieve efficiency. In this paper, we propose an unsupervised adversarial image retrieval (UAIR) framework by breaking the limitation of training labels. The framework is composed of two opposite parts and is linked by an adversarial loss function. For each input image, a generative model is used to select “well-matched” images from the database; a discriminative model is used to distinguish whether the selected images are similar enough to the input image. During training, the generative model tries to convince the discriminative model that the selected images are similar and the discriminative model always challenges the results of the generative model. The performances of the UAIR have been compared with other state-of-the-art image retrieval methods, including recently reported GAN-based methods. Extensive experiments show that the UAIR achieves significant improvement in CBIR with unsupervised adversarial training.

In this paper, a methodology for modelling two-phase flows based on a conservative level set method in the framework of finite volume method is presented. The novelty of the interface capturing method used here lies on the advection of level set which is solved with a WENO scheme and is corrected with a novel re-initialisation method for retaining its signed distance function character. The coupling with the volume of fluid method is done with a simple algebraic approach and with the new algorithm the accumulated mass conservation errors remain reasonably low. The paper presents a unique coupling between the level set method and the Eulerian-Lagrangian-Spray-Atomisation approach for modelling spray dispersion in liquid atomisation systems. The method is shown to have good accuracy providing similar results to other numerical codes for the classical tests presented. Preliminary results are also shown for three-dimensional simulations of the primary break-up of a turbulent liquid jet obtaining results comparable to direct numerical simulations. Consequently, the coupled method can be used for simulating various two-phase flow applications offering an accurate representation of the interface dynamics.

Off-lattice models are a well-established approach in multicellular modeling, where cells are represented as points that are free to move in space. The representation of cells as point objects is useful in a wide range of settings, particularly when large populations are involved; however, a purely point-based representation is not naturally equipped to deal with objects that have length, such as cell boundaries or external membranes. Here we introduce an off-lattice modeling framework that exploits rigid body mechanics to represent objects using a collection of conjoined one-dimensional edges in a viscosity-dominated system. This framework can be used to represent cells as free moving polygons, to allow epithelial layers to smoothly interact with themselves, to model rod-shaped cells such as bacteria and to robustly represent membranes. We demonstrate that this approach offers solutions to the problems that limit the scope of current off-lattice multicellular models. The authors present a framework for modeling cell interactions using rigid bodies, which can used to represent cells as free moving polygons, to allow epithelial layers to smoothly interact, to model bacteria and to robustly represent membranes.

Multimodal fusion is the idea of combining information in a joint representation of multiple modalities. The goal of multimodal fusion is to improve the accuracy of results from classification or regression tasks. This paper proposes a Bimodal Variational Autoencoder (BiVAE) model for audiovisual features fusion. Reliance on audiovisual signals in a speech recognition task increases the recognition accuracy, especially when an audio signal is corrupted. The BiVAE model is trained and validated on the CUAVE dataset. Three classifiers have evaluated the fused audiovisual features: Long-short Term Memory, Deep Neural Network, and Support Vector Machine. The experiment involves the evaluation of the fused features in the case of whether two modalities are available or there is only one modality available (i.e., cross-modality). The experimental results display the superiority of the proposed model (BiVAE) of audiovisual features fusion over the state-of-the-art models by an average accuracy difference $$\simeq $$ 3.28% and 13.28% for clean and noisy, respectively. Additionally, BiVAE outperforms the state-of-the-art models in the case of cross-modality by an accuracy difference $$\simeq $$ 2.79% when the only audio signal is available and 1.88% when the only video signal is available. Furthermore, SVM satisfies the best recognition accuracy compared with other classifiers.

An alternative power flow formulation, called linear power flow or “dc” power flow (DCPF), consists in the consideration of some simplifications regarding the network representation, leading to a linear problem formulation of low-cost computational performance, and reasonable accuracy. The distribution factors developed from the linear formulation of the power flow problem represent a sensitivity relation between system variables and have a variety of applications. Usually, the power transfer distribution factors (PTDF) are calculated based on the fact that all the power unbalance will be absorbed by the system’s reference bus. This paper proposes a new formulation of the PTDF to perform a more realistic analysis of a system’s operation, based on the linear power flow methodology and the generator’s participation factors. This formulation is then applied to an optimization problem to maximize wind power penetration in an 11-bus test system and the IEEE 118-bus system. The results are compared to those obtained by the original PTDF formulation to determine the new approach benefits. The results presented validate and indicate the effectiveness of the proposed technique.

This paper proposes a new representation for the Public Help Index θ (briefly, PHI θ). Based on winning coalitions, the PHI θ index was introduced by Bertini et al. in (2008). The goal of this article is to reformulate the PHI θ index using null player free winning coalitions. The set of these coalitions unequivocally defines a simple game. Expressing the PHI θ index by the winning coalitions that do not contain null players allows us in a transparent way to show the parts of the power assigned to null and non-null players in a simple game. Moreover, this new representation may imply a reduction of computational cost (in the sense of space complexity) in algorithms to compute the PHI θ index if at least one of the players is a null player. We also discuss some relationships among the Holler index, the PHI θ index, and the gnp index (based on null player free winning coalitions) proposed by Álvarez-Mozos et al. in (2015).

[EN[ In this work, a new mathematical algorithm for sparse and orthogonal constrained biplots, called CenetBiplots, is proposed. Biplots provide a joint representation of observations and variables of a multidimensional matrix in the same reference system. In this subspace the relationships between them can be interpreted in terms of geometric elements. CenetBiplots projects a matrix onto a low-dimensional space generated simultaneously by sparse and orthogonal principal components. Sparsity is desired to select variables automatically, and orthogonality is necessary to keep the geometrical properties that ensure the biplots graphical interpretation. To this purpose, the present study focuses on two different objectives: 1) the extension of constrained singular value decomposition to incorporate an elastic net sparse constraint (CenetSVD), and 2) the implementation of CenetBiplots using CenetSVD. The usefulness of the proposed methodologies for analysing high-dimensional and low-dimensional matrices is shown. Our method is implemented in R software and available for download from https://github.com/ananieto/SparseCenetMA., Open Access funding provided thanks to the CRUE-CSIC agreement with Springer Nature. This work was not supported by any grant., Publicación en abierto financiada por el Consorcio de Bibliotecas Universitarias de Castilla y León (BUCLE), con cargo al Programa Operativo 2014ES16RFOP009 FEDER 2014-2020 DE CASTILLA Y LEÓN, Actuación:20007-CL - Apoyo Consorcio BUCLE

Probabilistic inductive logic programming (PILP) is a statistical relational learning technique which extends inductive logic programming by considering probabilistic data. The ability to use probabilities to represent uncertainty comes at the cost of an exponential evaluation time when composing theories to model the given problem. For this reason, PILP systems rely on various pruning strategies in order to reduce the search space. However, to the best of the authors’ knowledge, there has been no systematic analysis of the different pruning strategies, how they impact the search space and how they interact with one another. This work presents a unified representation for PILP pruning strategies which enables end-users to understand how these strategies work both individually and combined and to make an informed decision on which pruning strategies to select so as to best achieve their goals. The performance of pruning strategies is evaluated both time and quality-wise in two state-of-the-art PILP systems with datasets from three different domains. Besides analysing the performance of the pruning strategies, we also illustrate the utility of PILP in one of the application domains, which is a real-world application.

Heterogeneous graph representation learning is designed to learn meaningful representation vectors from heterogeneous networks in few dimensions to extract the structure and features of the attributes of these networks. The embedding vector is the basis of and crucial to complex network analysis, and can be used in such downstream tasks as classification, clustering, link prediction, and recommendation. Key issues in heterogeneous graph neural networks pertain to ways to define heterogeneous neighbors and ways to aggregate them. Although considerable research has been devoted to homogeneous and heterogeneous network representation, the effective combination of information on the network structure and the attributes of nodes, especially effective use of meta-paths containing specific semantic information, remains rare. Here a meta-path-based heterogeneous graph neural network model is proposed. The meta-path is applied to sample the heterogeneous neighbors of each node in the network, and aggregate features of the same types of nodes to form type-related embedding. A multi-head attention mechanism is then applied to aggregate information on neighbors of different types of nodes and the model is trained by reducing context loss. Experiments on classification, clustering, link prediction, and recommendation tasks verified the validity of this model, which significantly improved the results of baseline methods.

Aspect-level sentiment classification aims to identify the sentiment polarity of a specific aspect in a sentence. In recent years, many researchers have sought to explore aspect-specific representation via attention mechanisms. Although a remarkable improvement in aspect-level sentiment classification has been achieved, these methods still suffer from deteriorative performance in certain cases for multiple reasons. First, they adopt a coarse-grained attention mechanism, which may cause information loss when the contextual sentence is long or includes multiple sentiments, and the aspect contains multiple words. Second, they consider only the influence of contextual keywords on sentiment polarity, ignoring the importance of key phrases in a sentence. To address these issues, a phrase-aware neural network based on fine-grained attention,referred to as FAPN, is proposed. The FAPN employs a convolutional neural network to extract phrase representations in the context and concatenates the representations with the corresponding word vector as input. Additionally, a fine-grained attention module is designed to generate aspect-specific representations by capturing the word-level interactions between the aspect and the sentence. Extensive experiments on five widely used benchmark datasets demonstrate the effectiveness of the proposed FAPN method.

Fourier series (FS) is an efficient tool for describing irregular geometries and has been employed to develop the FS-based particle model in the discrete element method (DEM). This work is devoted to extending the previous FS-based particle model to the Minkowski and Gilbert–Johnson–Keerthi (GJK)-based contact detection and resolution framework for DEM, and thus to improving its computational efficiency and compatibility with other conventional particle models. In the new FS-based particle model, instead of representing particle surface, the FS is proposed to represent the support function of particle surface. Particle surface and support points are then formulated based on the FS support function. As the Minkowski- and GJK-based detection and resolution framework relies heavily on the convexity of particles, the convexity constraint and the approach to generate convexity preserving FS-based particles are also presented. The accuracy of the new FS-based particle model for shape representation is analyzed using a set of irregular shape templates. DEM simulation of random packing and biaxial compression test with various particle models is also performed to demonstrate the computational performance and numerical stability of the new FS-based particle model.

The notable characteristic of large-scale linguistic decision-making problems is that there are so many decision makers who provide linguistic assessments by using fuzzy linguistic representation models. In real-world applications, fuzzy linguistic terms mean different things for different people, and linguistic assessments based on different linguistic representation models may be simultaneous in the same large-scale linguistic decision-making problems. To this end, a novel linguistic decision-making method based on the voting model is proposed in the paper to deal with multi-linguistic assessments provided by decision makers. In large-scale linguistic decision process, evaluation-based voting is defined and multi-linguistic decision matrix is designed to represent multi-linguistic assessments provided by decision makers by using different linguistic representation models, and properties of the decision matrix are analyzed to show that linguistic assessments based on different linguistic representation models can be simultaneously represented. Based on multi-linguistic decision matrix, a new linguistic decision-making framework is developed to deal with large-scale linguistic decision-making problems with multi-linguistic assessments, in which normalization of multi-linguistic decision matrix and trust degrees of linguistic terms are contained, and more important, based on trust degrees of linguistic terms and 2-tuple fuzzy linguistic aggregation operators, an uniform fusion method of multi-linguistic assessments is proposed to aggregate multi-linguistic assessments of large-scale linguistic decision-making problems. Finally, user experiences of shared bikes, which are a large-scale linguistic decision-making problem in real-world applications, are employed to show the new decision-making framework and the uniform fusion method of multi-linguistic assessments, and furthermore, compared with existing linguistic decision-making methods analyzed in the example, it seems that multi-linguistic decision matrix and the uniform fusion method are useful and effective tools to deal with large-scale linguistic decision-making problems with multi-linguistic assessments.

In reality, time series subject to the internal/external influence are usually characterized by nonlinearity, uncertainty, and incompleteness. Therefore, how to model the features of time series in nondeterministic environments is still an open problem. In this article, a novel high-order intuitionistic fuzzy cognitive map (HIFCM) is proposed, where intuitionistic fuzzy set (IFS) is introduced into fuzzy cognitive maps with temporal high-order structure. By means of IFS, the ability of model for the representation of uncertainty can be effectively improved. In order to capture the fluctuation features of series data, variational mode decomposition is utilized to decompose time series into sequences of various frequencies, based on which fine feature structures on different scales can be obtained. Each concept of HIFCM corresponds to one decomposed sequence such that casual reasoning can be achieved among the obtained features in various frequencies of time series. All parameters are learned by the particle swarm optimization algorithm. Finally, the performance of the method is verified on the public datasets, and experimental results show the feasibility and effectiveness of the proposed method.

We propose a few-shot learning method for spatial regression. Although Gaussian processes (GPs), or kriging, have been successfully used for spatial regression, they require many observations in the target task to achieve a high predictive performance. Our model is trained using spatial datasets on various attributes in various regions, and predicts values on unseen attributes in unseen regions given a few observed data. With our model, a task representation is inferred from given small data using a neural network. Then, spatial values are predicted by neural networks with a GP framework, in which task-specific properties are controlled by the task representations. The GP framework allows us to analytically obtain predictions that are adapted to small data. By using the adapted predictions in the objective function, we can train our model efficiently and effectively so that the test predictive performance improves when adapted to newly given small data. In our experiments, we demonstrate that the proposed method achieves better predictive performance than existing meta-learning methods using spatial datasets.

Reconstructing 3D face from sparse points is an ill-posed problem. While there already exits available solutions addressing this problem, to our knowledge, we propose a better-performed approach which can robustly reconstruct fine 3D face shape. Our method includes two modules: face model establishment based on probabilistic principal component analysis (PPCA) trained in an unsupervised manner to learn transformation between landmarks and point cloud in their low-dimensional representation, and 3D face reconstruction based on learned relation between them to reconstruct fine face shape. Overall, our method considers the probability of face shape and learns more useful information of 3D face shape. We compare our method with 3 typical and state-of-the-art methods on 2 datasets and the effectiveness of our method is demonstrated generally. Further experiments on datasets with noise of different intensities show the stability of our method.

Gesture recognition is the statistical representation used to identify specific gestures of motion involving the hands, head, face, and/or body. The paper addresses Hand Gesture Recognition (HGR) using novel machine learning and deep learning approaches. Machine learning algorithms viz. Artificial Neural Network (ANN) and Support Vector Machine (SVM) was implemented using spatial features comprising of geometrical features and Fourier descriptors. The experimental results revealed that ANN is better compared with SVM results. A novel decision fusion based two stream transfer learning approach was implemented using deep learning techniques in this paper. The two pretrained models MobileNet (0.5 and 1.0) and Inception V3 were used. The Region Of Interest (ROI) provided by YOLO was used to make a decision of which model is used for testing. The novel approach provided real time 99.6% recognition accuracy.

Conditioning complex subsurface flow models on nonlinear data is complicated by the need to preserve the expected geological connectivity patterns to maintain solution plausibility. Generative adversarial networks (GANs) have recently been proposed as a promising approach for low-dimensional representation of complex high-dimensional images. The method has also been adopted for low-rank parameterization of complex geologic models to facilitate uncertainty quantification workflows. A difficulty in adopting these methods for subsurface flow modeling is the complexity associated with nonlinear flow data conditioning. While conditional GAN (CGAN) can condition simulated images on labels, application to subsurface problems requires efficient conditioning workflows for nonlinear data, which is far more complex. We present two approaches for generating flow-conditioned models with complex spatial patterns using GAN. The first method is through conditional GAN, whereby a production response label is used as an auxiliary input during the training stage of GAN. The production label is derived from clustering of the flow responses of the prior model realizations (i.e., training data). The underlying assumption of this approach is that GAN can learn the association between the spatial features corresponding to the production responses within each cluster. An alternative method is to use a subset of samples from the training data that are within a certain distance from the observed flow responses and use them as training data within GAN to generate new model realizations. In this case, GAN is not required to learn the nonlinear relation between production responses and spatial patterns. Instead, it is tasked to learn the patterns in the selected realizations that provide a close match to the observed data. The conditional low-dimensional parameterization for complex geologic models with diverse spatial features (i.e., when multiple geologic scenarios are plausible) performed by GAN allows for exploring the spatial variability in the conditional realizations, which can be critical for decision-making. We present and discuss the important properties of GAN for data conditioning using several examples with increasing complexity.

In this paper, we show that the matrix elements of some Lorentz group representation operators and bases transform operators acting in the representation space may be expressed in terms of the modified Bessel–Clifford functions and their multi-index analogs introduced by the authors via Delerue hyper Bessel functions.

Recurrent Neural Networks (RNN) and their variants like Gated Recurrent Units (GRUs) have been the de-facto method in Natural Language Processing (NLP) for solving a range of NLP problems, including extractive text summarization. However, for certain sequential data with multiple temporal dependencies like the human text data, using a single RNN over the whole sequence might prove to be inadequate. Transformer models that use multiheaded attention have shown that human text contains multiple dependencies. Supporting networks like attention layers are needed to augment the RNNs to capture the numerous dependencies in text. In this work, we propose a novel combination of RNNs, called Parallel RNNs (PRNN), where small and narrow RNN units work on a sequence, in parallel and independent of each other, for the task of extractive text summarization. These PRNNs, without the need for any attention layers, capture various dependencies present in the sentence and document sequences. Our model achieved a 10% gain in ROUGE-2 score over the single RNN model on the popular CNN/Dailymail dataset. The boost in performance indicates that such an ensemble arrangement of RNNs improves the performance compared to the standard single RNNs, which alludes to the fact that constituent units of the PRNN learn various input sequence dependencies. Hence, the sequence is represented better using the combined representation from the constituent RNNs.

Capturing dynamic changes of networks can greatly improve the representation ability of nodes, which leads to dynamic network embedding becoming a hot research topic nowadays. However, current work focus on the correlation information and the position information of nodes, while the valuable timestamp information of edges is ignored. The timestamp information of edges presents the revolution of dynamic networks, which is extremely important for the dynamic node influence evaluation. To solve the problems of the existing works, we propose a novel dynamic network embedding method with multiple sequences learnings (DEMS). DEMS uses node sequence learning and edge sequence learning simultaneously to preserve more information of node dynamics in the network embedding. Specifically, node sequence learning preserves the node position information, and edge sequence learning preserves the edge timestamp information. Self-Attention mechanism is used in both sequence learnings to preserve the correlation information. Experiments on seven real-world dynamic networks verify the superiority of DEMS to the state-of-the-art methods in temporal link prediction tasks.

Recently, pose estimation algorithms have been widely discussed. Although been investigated by many prior works, pose estimation for heavily stacked workpieces, like bins, is still a challenge for industrial applications. Moreover, the scene of stacked workpieces is much more arduous than the general scene for robots to carry out routine tasks, such as object detection and picking. This paper aims to address the problem of pose estimation for stacked symmetrical workpieces in the presence of partial occlusions and cluttered backgrounds. To tackle those problems, we propose a novel pose estimation method for industrial stacking scenes based on RGB images. Specifically, due to the common symmetry characteristics of industrial workpieces, we firstly present a new standardized spatial representation method, which could auto-encode the 2D-3D correspondences of symmetrical workpieces. Besides, we introduce a novel GAN-based deep neural network model to reconstruct the representation of stacked workpieces. Based on that, the pose of the target workpieces is predicted based on the reconstructed expression and an improved RANSAC-PnP algorithm. Finally, comprehensive experiments demonstrate that the proposed method outperforms state-of-the-art methods, especially in complex stacking scenes.

With the increasing demand for secure biometric identification systems, finger vein recognition has received widespread attention. Recent studies have made progress in the verification of finger veins, but the extraction of the discriminative features of finger veins from images remains challenging. Although the traditional method of extracting features by combining softmax function and cross-entropy loss function can achieve separation between classes, it lacks discriminability. In addition, the method of setting a fixed margin yields good results, but it may cause some features to overlap. We argue that when some features are separated well from other features, the margin set needs to be reduced. Therefore, a dynamic margin softmax loss (dynamic softmax) is proposed in this study to obtain discriminative image features. Features and weight vectors are normalized, and the loss function dynamics are subsequently adjusted to achieve different cosine intervals for different classes. The main idea of this method is to maximize the distance between inter-class and minimize the distance between intra-class. This method can keep features separated without increasing the complexity of optimizing the neural network model. It is simpler and more effective than other loss functions. Experiments prove the effectiveness of the proposed method for finger vein recognition.

The goal of this study is to develop a prediction method to recognize the wake field behind a ship using a convolutional neural network (CNN) model. First, a new representation method for a 3D curved surface is proposed suitable for the CNN, called an image-based hull form representation (IHR). The advantages of the proposed method are the high fidelity of its hull form representation using more than 20,000 input data points and its fast prediction speed, which requires less than 0.01 s for a task that traditionally took more than an hour to estimate by physics-based simulation. The IHR regards that a two-dimensional grid formed on the 3D curved hull surface, which is used for structured-grid-based CFD, as a data set with the same data structure as the image data. Because CNNs recognize image data at accuracy rates higher than humans, a CNN is also be expected to recognize 3D surface characteristics with higher accuracy than humans. The image data are represented by three primary colors (cyan, magenta, yellow) in vertical and horizontal (i × j) pixels. The hull-form-structured grid can also be expressed as an i × j structure data with (x, y, z) coordinates that have the same data structure as the three primary colors in the image data. A CFD calculation data set of 2730 ship types with different stern shapes was constructed to verify the proposed method. The validation results proves that the root mean squared error of the proposed model is 0.005 to predict axial wake velocity on a propeller plane, and the coefficient of determination R2 achieves 0.986. In addition, the estimation speed for each hull-form prediction is 100,000 times faster than are physics-based simulations. The results lead to the conclusion that the representation method of a curved surface and the proposed prediction method of the stern wake field is a promising tool in the initial hull form design.

Multi-modal semantic trajectory prediction is of great importance for location-based applications. However, predicting trajectory is not trivial facing three challenges: (1) It is difficult to integrate useful information from multi-modal and heterogeneous data in different granularity for effective feature fusion; (2) All kinds of dependencies existing in multi-modal semantic trajectories are closely coupled and dynamically evolved, forming complex dependencies for which are difficult to quantify; (3) During the model training, the distribution of each modal feature shifts in different directions, resulting in the distortion of dependencies, which is accompanied by slow convergence and inaccurate predictions. In this paper, the Complex Dependencies Auto-learning Prediction Model (CDAPM) is proposed to solve these problems. First, the effective and robust representation of each points is obtained by jointly embedding multi-modal information. Then, the dependencies attention module is proposed to calculate the dependencies weight matrix and auto-learn the contribution of each point. Also, it solves the problem of long-term dependency effectively. Position Encoding and LSTM are used to enhance the time relationship of trajectory. Finally, Mode Normalization is designed to maintain prediction accuracy by preventing the distortion of dependencies and significantly accelerate the convergence speed. Experiments on two real data sets show that CDAPM outperforms the state-of-the-art methods.

We address the problem of unsupervised domain adaptation under the setting of generalized target shift (joint class-conditional and label shifts). For this framework, we theoretically show that, for good generalization, it is necessary to learn a latent representation in which both marginals and class-conditional distributions are aligned across domains. For this sake, we propose a learning problem that minimizes importance weighted loss in the source domain and a Wasserstein distance between weighted marginals. For a proper weighting, we provide an estimator of target label proportion by blending mixture estimation and optimal matching by optimal transport. This estimation comes with theoretical guarantees of correctness under mild assumptions. Our experimental results show that our method performs better on average than competitors across a range domain adaptation problems including digits,VisDA and Office. Code for this paper is available at https://github.com/arakotom/mars_domain_adaptation .

Click-through rate (CTR) prediction plays an important role in many industrial applications, feature engineering directly influences CTR prediction performance because features are normally the multi-field type. However, the existing CTR prediction techniques either neglect the importance of each feature or regard the feature interactions equally for feature learning. In addition, using an inner product or a Hadamard product is too simple to effectively model the feature interactions. These limitations lead to suboptimal performances of existing models. In this paper, we propose a framework called Hierarchical Attention and Feature Projection neural network (HAFP) for CTR prediction, which enables the automatically learning of more representative and efficient feature representation in an end-to-end manner. Towards this end, we employ a feature learning layer with a hierarchical attention mechanism to jointly extract more generalized and dominant features and feature interactions. In addition, a projective bilinear function is designed in meaningful second-order interaction encoder to effectively learn more fine-grained and comprehensive second-order feature interactions. Taking advantages of the hierarchical attention mechanism and the projective bilinear function, our proposed model can not only model feature learning in a flexible fashion, but also provide an interpretable capability of the prediction results. Experimental results on two real-world datasets demonstrate that HAFP outperforms the state-of-the-art in terms of Logloss and AUC for CTR prediction baselines. Further analysis verifies the importance of the proposed hierarchical attention mechanism and the projective bilinear function for modelling the feature representation, showing the rationality and effectiveness of HAFP.

Classifying 3D point clouds is an important and challenging task in computer vision. Currently, classification methods using multiple views lose characteristic or detail information during the representation or processing of views. For this reason, we propose a multi-view attention-convolution pooling network framework for 3D point cloud classification tasks. This framework uses Res2Net to extract the features from multiple 2D views. Our attention-convolution pooling method finds more useful information in the input data related to the current output, effectively solving the problem of feature information loss caused by feature representation and the detail information loss during dimensionality reduction. Finally, we obtain the probability distribution of the model to be classified using a full connection layer and the softmax function. The experimental results show that our framework achieves higher classification accuracy and better performance than other contemporary methods using the ModelNet40 dataset.

Clustering of face images is a well-known clustering problem with several applications as for example the automated face tagging in photo albums. The performance of face clustering algorithms largely depends on the face representations that are used. Therefore, the exploitation of the internal face representation (embedding) obtained by feeding the face image to a deep face network has led to increased clustering performance. In this work, we perform face clustering by combining multiple representations of each image obtained from several deep face networks. The multiple deep representations are not exploited in a straightforward manner (simple vector concatenation). Instead a weight is assigned to each representation to reflect its importance on the clustering result and a weighted clustering algorithm is employed to automatically adjust the weights and, consequently, the contribution of each representation on the clustering solution. It should also be noted that in the proposed approach, the number of clusters (number of faces) is not given as input, but it is automatically estimated using the silhouette criterion. The conducted experiments have shown that the weighted combination of deep face representations leads to improved clustering performance compared to using single representations.

High-performance segmentation of medical images, though important for automatic clinical diagnosis, remains a challenging problem. In this paper, we propose a novel polar regression network (PRNet) for the segmentation of oval-shaped targets in medical images. Through polar representation, the segmentation problem is formulated and decomposed into two sub-problems of center map estimation and ray length regression. The center map estimation is supervised by the probability center map, which is pre-generated by the contour-based algorithm. The ray length regression further leverages the center-attention polar loss and projection distance loss, in order to emphasize pixels in high probability in the center mask and capitalize on the implicit shape information. Experimental results demonstrate the effectiveness of our approach on the segmentation of oval targets in medical images.

Transformers, the dominant architecture for natural language processing, have also recently attracted much attention from computational visual media researchers due to their capacity for long-range representation and high performance. Transformers are sequence-to-sequence models, which use a self-attention mechanism rather than the RNN sequential structure. Thus, such models can be trained in parallel and can represent global information. This study comprehensively surveys recent visual transformer works. We categorize them according to task scenario: backbone design, high-level vision, low-level vision and generation, and multimodal learning. Their key ideas are also analyzed. Differing from previous surveys, we mainly focus on visual transformer methods in low-level vision and generation. The latest works on backbone design are also reviewed in detail. For ease of understanding, we precisely describe the main contributions of the latest works in the form of tables. As well as giving quantitative comparisons, we also present image results for low-level vision and generation tasks. Computational costs and source code links for various important works are also given in this survey to assist further development.

Visual trackers have achieved a high-level performance from deep features, but many limitations remain. Online trackers suffer from low speed while using deep features for parameter updating, and deep trackers trained offline demonstrate data hunger. To meet these challenges, our work aims to mine the target representation capability of a pre-trained model and presents deep convolutional descriptor aggregation (DCDA) for visual tracking. Based on spatial and semantic priors, we propose an edge-aware selection (EAS) and a central-aware selection (CAS) method to aggregate the accuracy-aware and robustness-aware features. To make full use of the scene context, our method is derived from one-shot learning by designing a dedicated regression process that is capable of predicting discriminative model in a few iterations. By exploiting robustness feature aggregation, the accuracy feature aggregation, and the discriminative regression, our DCDA with Siamese tracking architecture not only enhances the target prediction capacity, but also achieves a low-cost reuse of the pre-trained model. Comprehensive experiments on OTB-100, VOT2016, VOT2017, VOT2020, NFS30, and NFS240 show that our DCDA tracker achieves state-of-the-art performance with a high running speed of 65 FPS. The source code and all the experimental results of this work will be made public at https://github.com/Gitlyz007/DCDA_Tracker .

Gating transformation demonstrates great potential in recent deep convolutional neural networks design, enriching the feature representation and alleviating noisy signals by modeling the inter-channel dependencies using learnable parameters. However, the utilization of scaling approaches to reduce the redundancy of the hand-crafted attention mechanism has rarely been investigated. This paper proposes a novel scaled gated convolution that enables attention-enhanced CNNs to overcome the paradox between performance and redundancy. Our scaled gated convolution is a simple and effective alternative compared with both vanilla convolution and attention-enhanced convolutions, which can be easily applied to modern CNNs in a plug-and-play manner. Exhaustive experiments demonstrate that stacking scaled gated convolutions in baselines can significantly improve the performance in a broad range of visual recognition tasks, including image recognition, object detection, instance segmentation, keypoint detection, and panoptic segmentation, while obtaining a better trade-off between performance and attentive redundancy.

This paper investigates dynamically consistent updates of weighted unanimity rules given by variational Bewley preferences (VBP), a class of incomplete preferences characterized by pairs of utility index over consequences and ambiguity index over priors. We show that these dynamically consistent updates are also VBP that preserves the same ranking over consequences and feature conditional ambiguity index obtained through the full Bayesian update of the unconditional one. Moreover, we study conditional forced-choice relations, which capture choices that a decision-maker must take after learning some event. Formally, they are monotone continuous weak orders. We show that any conditional forced-choice relation that preserves the ambiguity attitude of a given unconditional VBP has a variational representation with the same affine utility function over consequences and the ambiguity index of its corresponding dynamically consistent update. Thus, our result provides a novel foundation for full Bayesian updating of variational preferences.

As a core element of the future smart city, intelligent transportation system provides comprehensive traffic analysis and prediction capabilities for traffic management. Vehicle type recognition, one of the indispensable parts of intelligent transportation, plays an important role in vehicle supervision, intelligent driving, public safety, and other aspects. However, fine-grained model recognition faces great challenges due to the substantial inter-category difference such as the complex background and various angles, as well as insufficient training data. This paper proposes a novel approach driven by small dataset, which finds the discriminative components based on the Faster-RCNN. Individual classifiers are trained for each component. The contribution of each component to vehicle type recognition can be evaluated through pre-trained VGG network. Considering the mutual dependency of these components, the component features are aggregated. It is assumed that the composited component features of the same model are on the potential manifold in the high-dimensional space; the SVM classifier operated at the category level is trained in the optimal transportation domain of the component features, which can enhance the discrimination ability of the feature representation. The experimental results show that the recognition accuracy of the proposed model in the Stanford BMW-10 dataset is 84.81%, which is obviously better than other models.

In this paper, a new method for the cross-term-free Wigner–Ville distribution (WVD) is proposed. The proposed method is based on sliding eigenvalue decomposition (SEVD) and the WVD which has been termed as SEVD–WVD. The SEVD decomposes the multicomponent signal into a set of monocomponent signals, and then, the WVD of analytic monocomponents is added to obtain a cross-term-free time–frequency representation. We have applied the proposed technique on clean and noisy synthetic signals in order to verify its robustness. Moreover, the SEVD–WVD is applied on speech signal to show its suitability on real-world data. The proposed method has been compared with other cross-term reduction techniques for the WVD in the literature.

In this paper, we aim to improve the performance of single-image super-resolution (SISR) by designing a more effective feature extraction module and a better fusion scheme for integrating hierarchical features. Firstly, we propose a selective multi-scale module (SMsM) to adaptively aggregate multi-scale features via self-learned weights and thus extract more distinctive representation. Then, we design an attentive global feature fusion (AGFF) scheme to reduce the redundant information inside the extracted hierarchical features by employing a gate mechanism (in the form of group convolution) and adaptively re-calibrate the features with channel-wise attention weights before fusion. Stacked SMsMs and AGFF compose a novel network which is termed selective multi-scale network (SMsN). Extensive experimental results demonstrate that our SMsN model outperforms some state-of-the-art SISR methods in terms of accuracy and efficiency.

In this paper, we provide an extension for partially linear models (PLMs) to allow the errors to follow a flexible class of two-piece distributions based on the scale mixtures of normal (TP-SMN) family. The TP-SMN is a rich class of distributions that covers symmetric/asymmetric as well as lightly/heavily tailed distributions which can be used to model datasets with outlying and also atypical data. Using a suitable hierarchical representation of the TP-SMN family developed specifically for PLM, we derived an EM-type algorithm for iteratively computing maximum penalized likelihood estimates of the proposed model parameters. We examined the performance of the proposed PLM model and methodology using simulation studies and a real dataset to show the robust aspects of this model.

As a flexible and powerful method to resolve strategy conflicts, the graph model for conflict resolution has drawn much attention. In the graph model for conflict resolution, decision-makers need to provide their preference information for all possible scenarios. Most existing studies assumed that decision-makers adopt quantitative representation formats. However, in some real-life situations, decision-makers may tend to use qualitative assessments due to their cognitive expression habits. In addition, stakeholders involved in a graph model can be a group that is composed of a large number of participants. How to manage these participants’ inconsistent preference assessments is also a debatable issue. To fit these gaps, in this study, we propose a graph model for conflict resolution with linguistic preferences, and this model allows participants to use inconsistent assessments. To do this, we first construct a linguistic preference structure, with the necessary concepts being defined. Then, four stability definitions for both a two-decision-maker scenario and an n-decision-maker scenario are introduced. To illustrate the usefulness of the proposed model, an illustrative example regarding the Huawei conflict is provided.

Agent-based models (ABMs) and state-and-transition simulation models (STSMs) have proven useful for understanding processes underlying social-ecological systems and evaluating practical questions about how systems might respond to different scenarios. ABMs can simulate a variety of agents (autonomous units, such as wildlife or people); agent characteristics, decision-making, adaptive behavior, and mobility; and agent-environment interactions. STSMs are flexible and intuitive stochastic landscape models that can track scenarios and integrate diverse data. Both can be run spatially and track metrics of management success. Due to the complementarity of these approaches, we sought to couple them through a dynamic linkage and demonstrate the relevance of this advancement for modeling landscape processes and patterns. We developed analytical techniques and software tools to couple these modeling approaches using NetLogo, R, and the ST-Sim package for SyncroSim. We demonstrated the capabilities and value of this coupled approach through a proof-of-concept case study of bison-vegetation interactions in Badlands National Park. The coupled ABM-STSM: (1) streamlined handling of model inputs and outputs; (2) allowed representation of processes at multiple temporal scales; (3) minimized assumptions; and (4) generated spatial and temporal patterns that better reflected agent-environment interactions. These developments constitute a new approach for representing agent-environment feedbacks; modelers can now use output from an ABM to dictate landscape changes within an STSM that in turn influence agents. This facilitates experimentation across domains (agent and environment) and creation of more realistic and management-relevant projections, and opens new opportunities for communicating models and linking to other methods.

Representation learning for video is increasingly gaining attention in the field of computer vision. For instance, video prediction models enable activity and scene forecasting or vision-based planning and control. In this article, we investigate the combination of differentiable physics and spatial transformers in a deep action conditional video representation network. By this combination our model learns a physically interpretable latent representation and can identify physical parameters. We propose supervised and self-supervised learning methods for our architecture. In experiments, we consider simulated scenarios with pushing, sliding and colliding objects, for which we also analyze the observability of the physical properties. We demonstrate that our network can learn to encode images and identify physical properties like mass and friction from videos and action sequences. We evaluate the accuracy of our training methods, and demonstrate the ability of our method to predict future video frames from input images and actions.

Topic models have been widely used for learning the latent explainable representation of documents, but most of the existing approaches discover topics in a flat structure. In this study, we propose an effective hierarchical neural topic model with strong interpretability. Unlike the previous neural topic models, we explicitly model the dependency between layers of a network, and then combine latent variables of different layers to reconstruct documents. Utilizing this network structure, our model can extract a tree-shaped topic hierarchy with low redundancy and good explainability by exploiting dependency matrices. Furthermore, we introduce manifold regularization into the proposed method to improve the robustness of topic modeling. Experiments on real-world datasets validate that our model outperforms other topic models in several widely used metrics with much fewer computation costs.

Knowledge graph completion(KGC) has attracted increasing attention in recent years, aiming at complementing missing relationships between entities in a Knowledge Graph(KG). While the existing KGC approaches utilizing the knowledge within KG could only complement a very limited number of missing relations, more and more approaches tend to study the completion of the multi-relationship knowledge graph. However, the existing completion methods of multi-relation knowledge graph regard knowledge graph as an undirected graph, which ignores the directionality of knowledge graph, so that the potential characteristics of multi-relation cannot be learned. Besides, most algorithms fail to explore the local information of knowledge because they ignore the different importance of entity adjacencies. In this paper, we propose to use local information fusion to join the entity and its adjacency relation, to acquiring the multi-relation representation. In addition, we try to specify distinct weights to model the direction of the relationship and apply the attention mechanism between entity nodes to obtain local information between entity nodes. Experiments conducted on three benchmark datasets and a medical domain knowledge graph dataset that we collect demonstrate the effectiveness of the proposed framework.

A turbulent medium with eddies of different scales gives rise to fluctuations in the index of refraction during the process of wave propagation, which interferes with the original spatial relationship, phase relationship and optical path. The outputs of two-dimensional imaging systems suffer from anamorphosis brought about by this effect. Randomness, along with multiple types of degradation, make it a challenging task to analyse the reciprocal physical process. Here, we present a generative adversarial network (TSR-WGAN), which integrates temporal and spatial information embedded in the three-dimensional input to learn the representation of the residual between the observed and latent ideal data. Vision-friendly and credible sequences are produced without extra assumptions on the scale and strength of turbulence. The capability of TSR-WGAN is demonstrated through tests on our dataset, which contains 27,458 sequences with 411,870 frames of algorithm simulated data, physical simulated data and real data. TSR-WGAN exhibits promising visual quality and a deep understanding of the disparity between random perturbations and object movements. These preliminary results also shed light on the potential of deep learning to parse stochastic physical processes from particular perspectives and to solve complicated image reconstruction problems given limited data. Turbulent optical distortions in the atmosphere limit the ability of optical technologies such as laser communication and long-distance environmental monitoring. A new method using adversarial networks learns to counter the physical processes underlying the turbulence so that complex optical scenes can be reconstructed.

Exhausters are families of compact, convex sets which provide minmax or maxmin representations of positively homogeneous functions and they are efficient tools for the study of nonsmooth functions (Demyanov in Optimization 45:13–29, 1999). Upper and lower exhausters of positively homogeneous functions are employed to describe optimality conditions in geometric terms and also to find directions of steepest descent or ascent. Since an upper/lower exhauster may contain finitely or infinitely many compact convex sets, the problem of minimality and reduction of exhausters naturally arise. There are several approaches to reduce exhausters (Abbasov in J Glob Optim 74(4):737–751, 2019; Grzybowski et al. in J Glob Optim 46(4):589–601, 2010; Kucuk et al. in J Optim Theory Appl 165:693–707, 2015; Roshchina in J Optim Theory Appl 136:261–273, 2008; J Convex Anal 15(4):859–868, 2008). In this study, in the sense of inclusion-minimality, some reduction techniques for upper exhausters of positively homogeneous functions defined from $$\mathbb {R}^2$$ to $$\mathbb {R}$$ is proposed by means of a representation of support functions. These techniques have concrete geometric meanings and they form a basis for a necessary and sufficient condition for inclusion-minimality of exhausters. Some examples are presented to illustrate each reduction technique.

We focus on the one-example person re-identification (Re-ID) task, where each identity has only one labeled example along with many unlabeled examples. Since each identity has only one labeled example, the number of initialized label examples is small, and the body parts of person are not aligned due to changes in person pose and camera angle under the camera. Therefore, the distinguishing information of learning labeled and unlabeled examples is challenging. To overcome these problems, we propose an end-to-end multi-task training network for semi-supervised Re-ID. First, we impose a part segmentation (PS) constraint on feature maps, forcing a module to predict part labels from the feature maps and enhance alignment. Second, we carefully design the network named Multiple Branch Network (MBN). MBN is a multi-branch deep network architecture, which consisting of one branch for global feature representation and two branches for local feature representation, local feature representation that including horizontal stripes representation and PS representation, respectively. Finally, loss function fusion is designed to learn discriminative features for semi-supervised Re-ID. Specifically, the MBN model is optimized by mining the object classification loss, exclusive loss and PS loss simultaneously. We validate the effectiveness of our approach by demonstrating its superiority over the state-of-the-art methods on the standard benchmark datasets, including Market-1501, DukeMTMC-reID. Notably, the rank-1 accuracy of our method outperforms the state-of-the-art method by 15.9 points (absolute, i.e., 71.7% vs. 55.8%) on Market-1501 and 8.9 points on DukeMTMC-reID.

Self-attention-based encoder-decoder frameworks have drawn increasing attention in recent years. The self-attention mechanism generates contextual representations by attending to all tokens in the sentence. Despite improvements in performance, recent research argues that the self-attention mechanism tends to concentrate more on the global context with less emphasis on the contextual information available within the local neighbourhood of tokens. This work presents the Dual Contextual (DC) module, an extension of the conventional self-attention unit, to effectively leverage both the local and global contextual information. The goal is to further improve the sentence representation ability of the encoder and decoder subnetworks, thus enhancing the overall performance of the translation model. Experimental results on WMT’14 English-German (En$$\rightarrow $$ → De) and eight IWSLT translation tasks show that the DC module can further improve the translation performance of the Transformer model.

Despite the great success achieved by convolutional neural networks in addressing the raindrop removal problem, the still relatively blurry results call for better problem formulations and network architectures. In this paper, we revisited the rainy-to-clean translation networks and identified the issue of imbalanced distribution between raindrops and varied background scenes. None of the existing raindrop removal networks consider this underlying issue, thus resulting in the learned representation biased towards modeling raindrop regions while paying less attention to the important contextual regions. With the aim of learning a more powerful raindrop removal model, we propose learning a soft mask map explicitly for mitigating the imbalanced distribution problem. Specifically, a two stage network is designed with the first stage generating the soft masks, which helps learn a context-enhanced representation in the second stage. To better model the heterogeneously distributed raindrops, a multi-scale dense residual block is designed to construct the hierarchical rainy-to-clean image translation network. Comprehensive experimental results demonstrate the significant superiority of the proposed models over state-of-the-art methods.

In this research paper, we consider the subspace clustering problem which aims at finding a low-dimensional representation of a high-dimensional data set. In particular, our central focus is upon the least squares regression based on which we elaborate an adaptive weighted least squares regression for subspace clustering. Compared to the least squares regression, we consider the data locality to adaptively select relevant and close samples and discard irrelevant and faraway ones. Additionally, we impose a weight matrix on the representation errors to adaptively highlight the meaningful features and minimize the effect of redundant/noisy ones. Finally, we also add a non-negativity constraint on the representation coefficients to enhance the graph interpretability. These interesting properties allow to build up a more informative and quality graph, thereby yielding very promising clustering results. Extensive experiments on synthetic and real databases demonstrated that our clustering method achieves consistently optimal results, compared to multiple clustering methods.

Since the outbreak of the COVID-19 pandemic, computer vision researchers have been working on automatic identification of this disease using radiological images. The results achieved by automatic classification methods far exceed those of human specialists, with sensitivity as high as 100% being reported. However, prestigious radiology societies have stated that the use of this type of imaging alone is not recommended as a diagnostic method. According to some experts the patterns presented in these images are unspecific and subtle, overlapping with other viral pneumonias. This report seeks to evaluate the analysis the robustness and generalizability of different approaches using artificial intelligence, deep learning and computer vision to identify COVID-19 using chest X-rays images. We also seek to alert researchers and reviewers to the issue of "shortcut learning". Recommendations are presented to identify whether COVID-19 automatic classification models are being affected by shortcut learning. Firstly, papers using explainable artificial intelligence methods are reviewed. The results of applying external validation sets are evaluated to determine the generalizability of these methods. Finally, studies that apply traditional computer vision methods to perform the same task are considered. It is evident that using the whole chest X-Ray image or the bounding box of the lungs, the image regions that contribute most to the classification appear outside of the lung region, something that is not likely possible. In addition, although the investigations that evaluated their models on data sets external to the training set, the effectiveness of these models decreased significantly, it may provide a more realistic representation as how the model will perform in the clinic. The results indicate that, so far, the existing models often involve shortcut learning, which makes their use less appropriate in the clinical setting.