Search

Abstract: In scenes with collectively moving objects, to disregard the individual objects and take the entire group into consideration for motion characterization is a promising approach with wide application prospects. In contrast to studies on the segmentation of independently moving objects, our purpose is to construct a segmentation of these objects to characterize their motions at a macroscopic level. In general, the collectively moving objects in a group have very similar motion behavior with their neighbors and appear as a kind of global collective motion. This paper presents a joint segmentation approach for these collectively moving objects. In our model, we extract these macroscopic movement patterns based on optical flow field sequences. Specifically, a group of collectively moving objects correspond to a region where the optical flow field has high magnitude and high local direction coherence. As a result, our problem can be addressed by identifying these coherent optical flow field regions. The segmentation is performed through the minimization of a variational energy functional derived from the Bayes classification rule. Specifically, we use a bag-of-words model to generate a codebook as a collection of prototypical optical flow patterns, and the class-conditional probability density functions for different regions are determined based on these patterns. Finally, the minimization of our proposed energy functional results in the gradient descent evolution of segmentation boundaries which are implicitly represented through level sets. The application of our proposed approach is to segment and track multiple groups of collectively moving objects in a large variety of real-world scenes. [Copyright &y& Elsevier]

Abstract: In this paper, Adjusted Rand Index (ARI) is generalized to two new measures based on matrix comparison: (i) Adjusted Rand Index between a similarity matrix and a cluster partition (ARImp), to evaluate the consistency of a set of clustering solutions with their corresponding consensus matrix in a cluster ensemble, and (ii) Adjusted Rand Index between similarity matrices (ARImm), to evaluate the consistency between two similarity matrices. Desirable properties of ARI are preserved in the two new measures, and new properties are discussed. These properties include: (i) detection of uncorrelatedness; (ii) computation of ARImp/ARImm in a distributed environment; and (iii) characterization of the degree of uncertainty of a consensus matrix. All of these properties are investigated from both the perspectives of theoretical analysis and experimental validation. We have also performed a number of experiments to show the usefulness and effectiveness of the two proposed measures in practical applications. [Copyright &y& Elsevier]

Abstract: Given a dataset P represented by an n×m matrix (where n is the number of data points and m is the number of attributes), we study the effect of applying transformations to P and how this affects the performance of different ensemble algorithms. Specifically, a dataset P can be transformed into a new dataset by a set of transformation operators in the instance dimension, such as sub-sampling, super-sampling, noise injection, and so on, and a corresponding set of transformation operators in the attribute dimension. Based on these conventional transformation operators and , a general form of the transformation operator is proposed to represent different kinds of transformation operators. Then, two new data transformation operators, known respectively as probabilistic based data sampling operator and probabilistic based attribute sampling operator, are designed to generate new datasets in the ensemble. Next, three new random transformation operators are proposed, which include the random combination of transformation operators in the data dimension, in the attribute dimension, and in both dimensions respectively. Finally, a new cluster ensemble approach is proposed, which integrates the random combination of data transformation operators across different dimensions, a hybrid clustering technique, a confidence measure, and the normalized cut algorithm into the ensemble framework. The experiments show that (i) random combination of transformation operators across different dimensions outperforms most of the conventional data transformation operators for different kinds of datasets. (ii) The proposed cluster ensemble framework performs well on different datasets such as gene expression datasets and datasets in the UCI machine learning repository. [Copyright &y& Elsevier]

Abstract: Recently, researchers are focusing more on the study of support vector machine (SVM) due to its useful applications in a number of areas, such as pattern recognition, multimedia, image processing and bioinformatics. One of the main research issues is how to improve the efficiency of the original SVM model, while preventing any deterioration of the classification performance of the model. In this paper, we propose a modified SVM based on the properties of support vectors and a pruning strategy to preserve support vectors, while eliminating redundant training vectors at the same time. The experiments on real images show that (1) our proposed approach can reduce the number of input training vectors, while preserving the support vectors, which leads to a significant reduction in the computational cost while attaining similar levels of accuracy. (2)The approach also works well when applied to image segmentation. [ABSTRACT FROM AUTHOR]

Abstract: In this paper, a quantization-based clustering algorithm (QBCA) is proposed to cluster a large number of data points efficiently. Unlike previous clustering algorithms, QBCA places more emphasis on the computation time of the algorithm. Specifically, QBCA first assigns the data points to a set of histogram bins by a quantization function. Then, it determines the initial centers of the clusters according to this point distribution. Finally, QBCA performs clustering at the histogram bin level, rather than the data point level. We also propose two approaches to improve the performance of QBCA further: (i) a shrinking process is performed on the histogram bins to reduce the number of distance computations and (ii) a hierarchical structure is constructed to perform efficient indexing on the histogram bins. Finally, we analyze the performance of QBCA theoretically and experimentally and show that the approach: (1) can be easily implemented, (2) identifies the clusters effectively and (3) outperforms most of the current state-of-the-art clustering approaches in terms of efficiency. [Copyright &y& Elsevier]

Abstract: Due to the popularity of computer games and computer-animated movies, 3D models are fast becoming an important element in multimedia applications. In addition to the conventional polygonal representation for these models, the direct adoption of the original scanned 3D point set for model representation is recently gaining more and more attention due to the possibility of bypassing the time consuming mesh construction stage, and various approaches have been proposed for directly processing point-based models. In particular, the design of a simplification approach which can be directly applied to 3D point-based models to reduce their size is important for applications such as 3D model transmission and archival. Given a point-based 3D model which is defined by a point set () and a desired reduced number of output samples , the simplification approach finds a point set which (i) satisfies ( being the cardinality of ) and (ii) minimizes the difference of the corresponding surface (defined by ) and the original surface (defined by ). Although a number of previous approaches has been proposed for simplification, most of them (i) do not focus on point-based 3D models, (ii) do not consider efficiency, quality and generality together and (iii) do not consider the distribution of the output samples. In this paper, we propose an Adaptive Simplification Method (ASM) which is an efficient technique for simplifying point-based complex 3D models. Specifically, the ASM consists of three parts: a hierarchical cluster tree structure, the specification of simplification criteria and an optimization process. The ASM achieves a low computation time by clustering the points locally based on the preservation of geometric characteristics. We analyze the performance of the ASM and show that it outperforms most of the current state-of-the-art methods in terms of efficiency, quality and generality. [Copyright &y& Elsevier]

Abstract: Pareto dominance is one of the most basic concepts in multi-objective optimization. However, it is inefficient when the number of objectives is large because in this case it leads to an unmanageable number of Pareto solutions. In order to solve this problem, a new concept of logic dominance is defined by considering the number of improved objectives and the quantity of improvement simultaneously, where probabilistic logic is applied to measure the quantity of improvement. Based on logic dominance, the corresponding logic nondominated solution is defined as a feasible solution which is not dominated by other ones based on this new relationship, and it is proved that each logic nondominated solution is also a Pareto solution. Essentially, logic dominance is an extension of Pareto dominance. Since there are already several extensions for Pareto dominance, some comparisons are given in terms of numerical examples, which indicates that logic dominance is more efficient. As an application of logic dominance, a house choice problem with five objectives is considered. [Copyright &y& Elsevier]

Abstract: In this paper, we propose a novel 3D head model retrieval framework. Specifically, to facilitate better classification and retrieval, the original 3D head model representations are embedded into another kernel feature space in which kernel principal component analysis (kernel PCA) is then performed to search for the optimal basis representation. Based on the extracted nonlinear features, a hierarchical indexing structure for 3D model retrieval is constructed using the hierarchical self organizing map (HSOM). The proposed indexing structure clusters the database into a hierarchy so that head models are partitioned by coarse features initially and then by finer scale features at lower levels. The main motivation of adopting this approach is that subspace technique like kernel PCA provides an elegant mechanism to describe the 3D head models on multiple resolutions based on the choices for reconstruction error and the orthogonal property of the produced eigenvectors. To further enhance the performance, a fuzzy metric between the query and the feature vector associated with each node on the SOMs is adopted instead of the usual Euclidean metric. Only nodes that possess high fuzzy measure values will be considered further for retrieval. In this way, the fuzzy measure approach is able to pick up potential relevant models even though they may be distributed across a number of neighbouring nodes. In addition to model categorization, the topology-preserving property of HSOM also facilitates the exploration of the model database with the possibility for further knowledge discovery. The effectiveness of the proposed approach is verified by a set of simulation examples on a 3D head model database. [Copyright &y& Elsevier]

Abstract: In this paper, we address the problem of extracting gene regulation information from microarray data for cancer classification. From the biological viewpoint, a model of gene regulation probability is established where three types of gene regulation states in a tissue sample are assumed and then two regulation events correlated with the class distinction are defined. Different from the previous approaches, the proposed algorithm uses gene regulation probabilities as carriers of regulation information to select genes and construct classifiers. The proposed approach is successfully applied to two public available microarray data sets, the leukemia data and the prostate data. Experimental results suggest that gene selection based on regulation information can greatly improve cancer classification, and the classifier based on regulation information is more efficient and more stable than several previous classification algorithms. [Copyright &y& Elsevier]

Abstract: Developed from the conventional Fourier transform, the fractional Fourier transform is a powerful signal analysis and processing technique. In this paper, we apply it to the field of face recognition. By combining it with the discrimination analysis technique, we propose a new face recognition approach. First, we use a two-dimensional separability judgment to select appropriate value of angle parameter for discrete fractional Fourier transform. Second, we present a reformative Fisherface method to extract discriminative features from the preprocessed images and perform the classification using the nearest neighbor classifier. Experimental results on two public face databases indicate that our approach outperforms four representative discrimination methods. It obtains better and robust classification effects. [Copyright &y& Elsevier]

Abstract: Two-dimensional (2D) discrimination analysis using methods such as 2D PCA and Image LDA is of interest in face recognition because it extracts discriminative features faster than one-dimensional (1D) discrimination analysis. However, existing 2D methods generally use more discriminative features and take longer to test than 1D methods. 2D PCA in particular cannot make full use of the Fisher discriminant criterion. Image LDA also has drawbacks in that it cannot perform 2D principal component analysis and discards components with poor discriminative capabilities. In addition, existing 2D methods cannot provide an automatic strategy to choose 2D principal components or discriminant vectors. In this paper, we propose 2D Fisherface, a novel discrimination approach that combines the two-stage “” strategy and 2D discrimination techniques. It can extract face discriminative features by automatically selecting two-dimensional principal components and discriminant vectors. Using the AR database as the test data, it is shown that the proposed approach is faster and more effective than several representative 1D and 2D discrimination methods. [Copyright &y& Elsevier]

Abstract: This paper introduces a novel framework for 3D head model recognition based on the recently proposed 2D subspace analysis method. Two main contributions have been made. First, a 2D version of clustering-based discriminant analysis (CDA) is proposed, which combines the capability to model the multiple cluster structure embedded within a single class with the computational advantage that is characteristic of 2D subspace analysis methods. Second, we extend the applications of 2D subspace methods to the field of 3D head model classification by characterizing these models with 2D feature sets. [Copyright &y& Elsevier]

Abstract: The Fisherface method is the most representative method of the linear discrimination analysis (LDA) technique. However, there persists in the Fisherface method at least two areas of weakness. The first weakness is that it cannot make the achieved discrimination vectors completely satisfy the statistical uncorrelation while costing a minimum of computing time. The second weakness is that not all the discrimination vectors are useful in pattern classification. In this paper, we propose an uncorrelated Fisherface approach (UFA) to improve the Fisherface method in these two areas. Experimental results on different image databases demonstrate that UFA outperforms the Fisherface method and the uncorrelated optimal discrimination vectors (UODV) method. [Copyright &y& Elsevier]

Classification of 3D head models based on their shape attributes for subsequent indexing and retrieval are important in many applications, as in hierarchical content-based retrieval of these head models for virtual scene composition, and the automatic annotation of these characters in such scenes. While simple feature representations are preferred for more efficient classification operations, these features may not be adequate for distinguishing between the subtly different head model classes. In view of these, we propose an optimization approach based on genetic algorithm (GA) where the original model representation is transformed in such a way that the classification rate is significantly enhanced while retaining the efficiency and simplicity of the original representation. Specifically, based on the Extended Gaussian Image (EGI) representation for 3D models which summarizes the surface normal orientation statistics, we consider these orientations as random variables, and proceed to search for an optimal transformation for these variables based on genetic optimization. The resulting transformed distributions for these random variables are then used as the modified classifier inputs. Experiments have shown that the optimized transformation results in a significant improvement in classification results for a large variety of class structures. More importantly, the transformation can be indirectly realized by bin removal and bin count merging in the original histogram, thus retaining the advantage of the original EGI representation. [Copyright &y& Elsevier]

In this paper, we propose a novel corpus-based enrichment approach for short text clustering. Since sparseness brings about the problem of insufficient word co-occurrence and lack of context information, previous researches use external sources such as Wikipedia or WordNet to enrich the representation of short text documents, which requires extra resources and might lead to possible inconsistency. On the other hand, corpus-based approaches use no external information in mining short text data. By introducing a set of conjugate definitions to characterize the structures of topics and words, and by proposing a virtual generative procedure for short texts, we perform expansion on short text data. Specifically, new words which may not appear in a short text document were added with a virtual term frequency, and this virtual frequency is obtained from the posterior probabilities of new words given all the words in that document. The complete procedure can be regarded as mapping data points (documents) from the original feature space to a hidden semantic space (topic space). After performing semantic smoothing, data points are then mapped back to the original space. We conduct experiments on two short text datasets, and the results show that the proposed method can effectively address the sparseness problem. For these datasets, our method, using only a basic clustering algorithm, attains a comparable performance with methods based on enrichment with external information sources. [ABSTRACT FROM AUTHOR]

Sufficient labeled training data may not be available for pedestrian detection in many real-world scenes. Semi-supervised settings naturally apply for the case where an adequate number of images are collected in a target scene but only a small proportion of them can be manually annotated. A common strategy is to adopt a detector trained on a well-established dataset (source data) or the limited annotated data to pseudo-annotate unannotated images. However, the domain gap and the lack of supervision in the target scene may lead to low-quality pseudo annotations. In this paper, we propose a Scene-adaptive Pseudo Annotation (SaPA) approach, which aims at exploiting two types of training data: source data providing sufficient supervision and unannotated target data offering domain-specific information. To utilize the source data, an Annotation Network (AnnNet) competes with a domain discriminator to learn domain-invariant features. To exploit the unannotated data, we temporally aggregate the parameters of AnnNet to build a more robust network, which is able to provide training goals for AnnNet. This new approach improves the generalization performance of AnnNet, which eventually leads to high-quality pseudo annotations to the unannotated data. Both manual and pseudo annotations are leveraged to train a more precise and scene-specific detector. We perform extensive experiments on multiple benchmarks to verify the effectiveness and superiority of SaPA. [ABSTRACT FROM AUTHOR]

Domain adaptation has achieved significant progress recently by adapting models trained on a source domain to an unlabeled target domain. Open Set Domain adaptation (OSDA) has drawn much attention nowadays, where the target domain contains some exclusive categories other than the source domain's known classes. With no label in the target data, existing OSDA methods often suffer from negative transfer. Conventional methods for unknown class rejection require an empirical setting of the confidence threshold, which lacks flexibility since the model confidence may vary during the training process, and our motivation is to omit the effort of setting the rejection threshold manually. Based on the idea that latent features of the same class should be in the same cluster to address this issue, we propose a domain adaptive open set recognition framework: Cluster-based Adversarial Decision Boundary (CADB). Specifically, we design an end-to-end unknown class rejection model consisting of three components: known class prototype estimation under the cluster assumption; known class similarity score estimation; and adaptive unknown class rejection threshold generation with adversarial feature suppression. These three components work as one entity to give a similarity score for each sample. Those samples that are less similar to the cluster prototype compared with the counterfactual features are rejected as the unknown class. Extensive evaluations are conducted to verify the effectiveness and robustness of the proposed boundary generation procedure. [ABSTRACT FROM AUTHOR]

Abstract: Portfolio selection theory with fuzzy returns has been well developed and widely applied. Within the framework of credibility theory, several fuzzy portfolio selection models have been proposed such as mean–variance model, entropy optimization model, chance constrained programming model and so on. In order to solve these nonlinear optimization models, a hybrid intelligent algorithm is designed by integrating simulated annealing algorithm, neural network and fuzzy simulation techniques, where the neural network is used to approximate the expected value and variance for fuzzy returns and the fuzzy simulation is used to generate the training data for neural network. Since these models are used to be solved by genetic algorithm, some comparisons between the hybrid intelligent algorithm and genetic algorithm are given in terms of numerical examples, which imply that the hybrid intelligent algorithm is robust and more effective. In particular, it reduces the running time significantly for large size problems. [Copyright &y& Elsevier]

The principal component in the application of biometrics authentication (BA) technologies, the biometric data usually shows three characteristics: large numbers of individuals, small sample size and high dimensionality. One of the major research difficulties of BA is the single sample biometrics recognition problem. We often face this problem in real-world applications. It may lead to a bad recognition result. To solve this problem, we present a novel approach based on feature level biometrics fusion. We combine two kinds of biometrics: one is the face feature which is a representative of contactless biometrics, and another is the palmprint feature which is a typical contact biometrics. We extract the discriminant feature using Gabor-based image preprocessing and principal component analysis (PCA) techniques. And then design a distance-based separability weighting strategy to conduct feature level fusion. Using a large face database and a large palmprint database as the test data, the experimental results show that the presented approach significantly improves the recognition effect of single sample biometrics problem, and there is strong supplement between face and palmprint biometrics. [Copyright &y& Elsevier]

Abstract: In this paper, a kernelized version of clustering-based discriminant analysis is proposed that we name KCDA. The main idea is to first map the original data into another high-dimensional space, and then to perform clustering-based discriminant analysis in the feature space. Kernel fuzzy c-means algorithm is used to do clustering for each class. A group of tests on two UCI standard benchmarks have been carried out that prove our proposed method is very promising. [Copyright &y& Elsevier]

Intelligent traffic signal control helps to reduce traffic congestion and thus has been studied for a few decades. Multi-intersection cooperative traffic signal control (CTSC), which is more practical than single-intersection traffic signal control, has attracted much attention and research in recent years. Existing works on multi-intersection CTSC make responsive policies based on the sequence of agents' actions. One issue in multi-intersection CTSC is that every agent's actions are mapped from its own road information and some useful information, e.g., the distance of adjacent agents, is ignored, which may lead to suboptimal traffic signal control policies. To address this issue, in this paper a decentralized coordination graph algorithm, referred to as Multi-step return and Off-policy Asynchronous Advantage Actor-Critic Graph (MOA3CG) algorithm, is proposed. The MOA3CG algorithm is based on an asynchronous method of multiagent deep reinforcement learning and a coordination graph; the proposed algorithm makes traffic signal control policies based on current traffic states, the history of observations and other information. A new reward function and An Adjusting Matrix of Traffic Signal Phase Control (AMTSPC) are proposed, which are used by the MOA3CG algorithm in the policy-making process; the AMTSPC is to alter selection of actions by considering the distance of adjacent agents. Experimental results on real-world road scenarios show that the proposed algorithm outperforms other four state-of-the-art algorithms in terms of average delay, average traveling time of vehicles, and the throughput of vehicles, thus eventually helps to mitigate traffic congestion. • MOA3CG algorithm is proposed for multi-intersection cooperative traffic signal control. • The MOA3CG algorithm is based on multi-step return, off-policy and a graph. • An AMTSPC for selection of actions is proposed and adopted in MOA3CG. • Experiments show that MOA3CG algorithm outperforms the state-of-the-art algorithms. [ABSTRACT FROM AUTHOR]

Highlights • we develop a new formulation for MTL based on the decomposition of the coeffcient matrix into a Hadamard (element-wise) product of two matrices. Comparing with conventional MTL methods and EL, the advantages of the proposed approach can be summarized as follows: (1) It is capable of capturing the competitive structure among tasks. (2) Unimportant features which are common across the tasks can be removed from the final model. Moreover, we propose to employ an alternating optimization method to iteratively estimate the coeffcients of the two components in the SpEL objective function. • We also provide an analysis of the proposed model based on the element- wise product decomposition framework to highlight its advantage. • We conduct experimental studies on both synthetic and real data in different application domains which include handwritten digit data and gene expression analysis. The experimental results demonstrate the effectiveness of the proposed model, and suggest potential applications of the proposed method. Abstract Multi-task learning (MTL) aims to enhance generalization performance by exploring the inherent structures across tasks. Most existing MTL methods are based on the assumption that the tasks are positively correlated, and utilize the shared structures among tasks to improve learning performance. By contrast, there also exist competitive structure (negative relationships) among tasks in some real-world applications, and conventional MTL methods which explore shared structures across tasks may lead to unsatisfactory performance in this setting. Another challenge, especially in a high dimensional setting, is to exclude irrelevant features (sparse structure) from the final model. For this purpose, this work propose a new method, which is referred to as Sparse Exclusive Lasso (SpEL) for multi-task learning. The proposed SpEL is able to capture the competitive relationship among tasks (competitive structure), while remove unimportant features which are common across the tasks from the final model (sparse structure). Experimental studies on synthetic and real data indicate that the proposed method can significantly improve learning performance by identifying sparse and task-competitive structures simultaneously. [ABSTRACT FROM AUTHOR]

Highlights • Overlapping between feature groups of different clusters is allowed to encourage the shared information and distinct information simultaneously; • Representative Dual Features (RepDFs) is introduced to evaluate the correlations between clusters; • Hierarchical structural similarities between clusters are explored in feature space and sample space from the topological perspective; • Correlations from RepDFs are incorporated into hierarchical structural similarities to guide knowledge transfer, which increases diversities of instances by exploiting instances from related clusters. Abstract Unsupervised multi-task learning exploits the shared knowledge to improve performances by learning related tasks simultaneously. In this paper, we propose an unsupervised multi-task learning method with hierarchical data structure. It strengthens similarities between instances in the same cluster, and increases diversities of instances by utilizing instances from related clusters. Firstly, we introduce Rep resentative D ual F eature s (RepDFs) that possess representative capabilities in the feature space and the sample space for each cluster concurrently. Secondly, we explore hierarchical structural similarities between clusters in related tasks from the topological perspective: 1) feature basis matrix, which learns compact representations for features in the feature space; and 2) sample refined matrix, which preserves local structures in the sample space. Thirdly, we adopt RepDFs to measure correlations between clusters and incorporate hierarchical structural similarities to conduct knowledge transfer among tasks. Experimental results on real-world data sets demonstrate the effectiveness and superiority of the proposed method over existing multi-task clustering methods. [ABSTRACT FROM AUTHOR]

Activation functions play important roles in deep convolutional neural networks. This work focuses on learning activation functions via combining basic activation functions in a data-driven way. We explore three strategies to learn the activation functions, and allow the activation operation to be adaptive to inputs. We firstly explore two strategies to linearly and nonlinearly combine basic activation functions, respectively. Then we further investigate a strategy that basic activation functions are combined in a way of a hierarchical integration. Experiments demonstrate that the proposed activation functions lead to better performances than ReLU and its variants on benchmarks with various scales. [ABSTRACT FROM AUTHOR]

Generative Adversarial Networks (GANs) receive extensive attention due to its capability of high-fidelity image synthesis. The synthesized content can be edited by imposing transformations on the latent codes. To edit real images, the GAN inversion methods mainly focus on learning an inverse mapping from the data space to the latent space of a well-trained GAN through a separate encoder. In this paper, we propose a Discriminator Feature-based Progressive Inversion (DFPI) model for GAN-based image reconstruction and enhancement. We find that the expressiveness of the generator feature is richer than that of the latent code for a given input image. To achieve high-quality reconstruction, we estimate the generator features in a light-weight optimization process, conditioned on off-the-shelf discriminator features. Specifically, a recomposition matrix is learnt to map high-level discriminator features to the high-level generator features due to their reverse structure. To progressively improve the reconstruction quality, we adopt a convolutional correction module to alleviate the estimation error at a lower generator layer, conditioned on the features of a lower discriminator layer. By performing the progressive correction process, the input images can be accurately reconstructed. Note that the resulting recomposition matrix and correction module can be generalized to unseen images. The experiments demonstrate the superior performance of DFPI over competing methods on GAN inversion as well as a variety of image enhancement tasks. [ABSTRACT FROM AUTHOR]

• Transferring the training experience among constituent networks facilitates semi-supervised image classification. • We design a collaborative learning mechanism based on unreliability adaptation among constituent networks. • We improve the complementarity of constituent networks by resisting adversarial perturbation. • The unreliability adaptation and perturbation-based regularization lead to the superior performance on multiple datasets. Constructing training goals for unlabeled data is crucial for image classification in the semi-supervised setting. Consistency regularization typically encourages a model to produce consistent predictions with the given training goals, while unreliability adaptation aims to learn the transition probabilities from model predictions to training goals, instead of enforcing their consistency. In this paper, we present a model of Collaborative learning with Unreliability Adaptation (CoUA), in which multiple constituent networks collaboratively learn with each other by adapting their predictions. Toward this end, an additional adaptation module is incorporated into each network to learn a transition probability from its own prediction to that of the paired network. Therefore, the networks can exchange training experience, without being overly sensitive to the unreliability of predictions. To further enhance the collaborative learning, each network is encouraged to produce consistent predictions with the consensus results, while being resistant to the adversarial perturbations against others. Therefore, the networks are able to mutually reinforce each other. We perform extensive experiments on multiple image classification benchmarks to verify the superiority of the co-adaptation based collaborative learning mechanism. [ABSTRACT FROM AUTHOR]

There are not many classifier ensemble approaches which investigate the data sample space and the feature space at the same time, and this multi-pronged approach will be helpful for constructing more powerful learning models. For example, the AdaBoost approach only investigates the data sample space, while the random subspace technique only focuses on the feature space. To address this limitation, we propose the progressive subspace ensemble learning approach (PSEL) which takes into account the data sample space and the feature space at the same time. Specifically, PSEL first adopts the random subspace technique to generate a set of subspaces. Then, a progressive selection process based on new cost functions that incorporate current and long-term information to select the classifiers sequentially will be introduced. Finally, a weighted voting scheme is used to summarize the predicted labels and obtain the final result. We also adopt a number of non-parametric tests to compare PSEL and its competitors over multiple datasets. The results of the experiments show that PSEL works well on most of the real datasets, and outperforms a number of state-of-the-art classifier ensemble approaches. [ABSTRACT FROM AUTHOR]

Generative Adversarial Networks (GANs) have brought great progress in image-to-image translation. The problem that we focus on is how to accurately extract and transfer the makeup style from a reference facial image to a target face. We propose a GAN-based generative model with Target-aware makeup Style Encoding and Verification, which is referred to as TSEV-GAN. This design is due to the following two insights: (a) When directly encoding the reference image, the encoder may focus on regions which are not necessarily important or desirable. To precisely capture the style, we encode the difference map between the reference and corresponding de-makeup images, and then inject the obtained style code into a generator. (b) A generic real-fake discriminator cannot ensure the correctness of the rendered makeup pattern. In view of this, we impose style representation learning on a conditional discriminator. By identifying style consistency between the reference and synthesized images, the generator is induced to precisely replicate the desirable makeup. We perform extensive experiments on the existing makeup benchmarks to verify the effectiveness of our improvement strategies in transferring a variety of makeup styles. Moreover, the proposed model is able to outperform other existing state-of-the-art makeup transfer methods in terms of makeup similarity and irrelevant content preservation. [ABSTRACT FROM AUTHOR]

• We exploit the Siamese Network to find a clustering-friendly embedding space to mine highly-reliable pseudo-supervised information for the application of VAT and Conditional-GAN to synthesize cluster-specific samples in the setting of unsupervised learning. • We proposed adopting VAT to synthesize samples with different levels of perturbations that can enhance the robustness of Feature Extractor to noise and improve the lower-dimensional latent coding space discovered by the Feature Extractor. • We conducted experiments to verify that the latent space discovered by the Feature Extractor can facilitate the Siamese Network to find a clustering-friendly embedding space and extract pseudo-supervised information for VAT and Conditional-GAN. • The training of our EDCN involves the adversarial gaming between three players, which not only boosts performance improvement of the clustering but also preserves the cluster-specific information from the Siamese Network in synthesizing samples. In this paper, we propose an enhanced deep clustering network (EDCN), which is composed of a Feature Extractor, a Conditional Generator, a Discriminator and a Siamese Network. Specifically, we will utilize two kinds of generated data based on adversarial training, as well as the original data, to train the Feature Extractor for learning effective latent representations. In addition, we adopt the Siamese network to find an embedding space, where a better affinity similarity matrix is obtained as the key to success of spectral clustering in providing reliable pseudo-labels. Particularly, the obtained pseudo-labels will be used to generate realistic data by the Generator. Finally, the discriminator is used to model the real joint distribution of data and corresponding latent representations for Feature Extractor enhancement. To evaluate our proposed EDCN, we conduct extensive experiments on multiple data sets including MNIST, USPS, FRGC, CIFAR-10, STL-10, and Fashion-MNIST by comparing our method with a number of state-of-the-art deep clustering methods, and experimental results demonstrate its effectiveness and superiority. [ABSTRACT FROM AUTHOR]

Data augmentation via randomly combining training instances and interpolating the corresponding labels has shown impressive gains in image classification. However, model attention regions are not necessarily meaningful in class semantics, especially for the case of limited supervision. In this paper, we present a semi-supervised classification model based on Class-Ambiguous Data with Attention Regularization, which is referred to as CADAR. Specifically, we adopt a Random Regional Interpolation (RRI) module to construct complex and effective class-ambiguous data, such that the model behavior can be regularized around decision boundaries. By aggregating the parameters of a classification network over training epochs to produce more reliable predictions on unlabeled data, RRI can also be applied to them as well as labeled data. Further, the classifier is enforced to apply consistent attention on the original and constructed data. This is important for inducing the model to learn discriminative features from the class-related regions. The experiment results demonstrate that CADAR significantly benefits from the constructed data and attention regularization, and thus achieves superior performance across multiple standard benchmarks and different amounts of labeled data. [ABSTRACT FROM AUTHOR]

Abstract: Cluster ensemble approaches make use of a set of clustering solutions which are derived from different data sources to gain a more comprehensive and significant clustering result over conventional single clustering approaches. Unfortunately, not all the clustering solutions in the ensemble contribute to the final result. In this paper, we focus on the clustering solution selection strategy in the cluster ensemble, and propose to view clustering solutions as features such that suitable feature selection techniques can be used to perform clustering solution selection. Furthermore, a hybrid clustering solution selection strategy (HCSS) is designed based on a proposed weighting function, which combines several feature selection techniques for the refinement of clustering solutions in the ensemble. Finally, a new measure is designed to evaluate the effectiveness of clustering solution selection strategies. The experimental results on both UCI machine learning datasets and cancer gene expression profiles demonstrate that HCSS works well on most of the datasets, obtains more desirable final results, and outperforms most of the state-of-the-art clustering solution selection strategies. [Copyright &y& Elsevier]

Abstract: Antigenic peptides presented to T cells by MHC molecules are essential for T or B cells to proliferate and eventually differentiate into effector cells or memory cells. MHC binding prediction is an active research area. Reliable predictors are demanded to identify potential vaccine candidates. The recent kernel-based algorithm KernelRLSpan (Shen et al., 2013) shows promising power on MHC II binding prediction. Here, KernelRLSpan is modified and applied to MHC I binding prediction, which we refer to as KernelRLSpanI. Besides this, we develop a novel consensus method to predict naturally processed peptides through integrating KernelRLSpanI with two state-of-the-art predictors NetMHCpan and NetMHC. The consensus method achieved top performance in the Machine Learning in Immunology (MLI) 2012 Competition, 3 [3] URL: http://bio.dfci.harvard.edu/DFRMLI/HTML/natural.php. group 2. We also introduce our progress of improving our MHC II binding prediction method KernelRLSpan by diffusion map. [Copyright &y& Elsevier]

• We propose a residual generator (Rg-GAN) served as a better approximation divergence minimization framework for GAN, and prove that residual generator for standard and least-squares GAN are equivalent to the minimization of reverse-KL and a new instance of f-divergence, respectively. • We prove that Rg-GAN can be reduced to IPMs based GAN and bridge the gap between IPMs and f-divergence. • We propose a new loss function for the discriminator of Rg-GAN that manifests a better discriminative property and therefore improved on Rg-GAN generalisation ability. • We conduct experiments on multiple benchmark data sets and demonstrate that our proposed framework can mitigate the mode collapse issue and facilitate GAN to generate higher-quality images with negligible additional computation cost. GAN is a generative modelling framework which has been proven as able to minimise various types of divergence measures under an optimal discriminator. However, there is a gap between the loss function of GAN used in theory and in practice. In theory, the proof of the Jensen divergence minimisation involves the min-max criterion, but in practice the non-saturating criterion is instead used to avoid gradient vanishing. We argue that the formulation of divergence minimization via GAN is biased and may yield a poor convergence of the algorithm. In this paper, we propose the Residual Generator for GAN (Rg-GAN), which is inspired by the closed-loop control theory, to bridge the gap between theory and practice. Rg-GAN minimizes the residual between the loss of the generated data to be real and the loss of the generated data to be fake from the perspective of the discriminator. In this setting, the loss terms of the generator depend only on the generated data and therefore contribute to the optimisation of the model. We formulate the residual generator for standard GAN and least-squares GAN and show that they are equivalent to the minimisation of reverse-KL divergence and a novel instance of f-divergence, respectively. Furthermore, we prove that Rg-GAN can be reduced to Integral Probability Metrics (IPMs) GANs (e.g., Wasserstein GAN) and bridge the gap between IPMs and f-divergence. Additionally, we further improve on Rg-GAN by proposing a loss function for the discriminator that has a better discrimination ability. Experiments on synthetic and natural images data sets show that Rg-GAN is robust to mode collapse, and improves the generation quality of GAN in terms of FID and IS scores. [ABSTRACT FROM AUTHOR]

Low-resolution (LR) face verification has received much attention because of its wide applicability in real scenarios, especially in long-distance surveillance. However, the poor quality and scarcity of training data make the accuracy far from satisfactory. In this paper, we propose an end-to-end LR face translation and verification framework to improve the generation quality of face images and face verification accuracy simultaneously. We design a dual domain adaptive structure to generate high-quality images. On one hand, the structure can reduce the domain gap between training data and test data. On the other hand, the structure preserves identity consistency and low-level attributes. Meanwhile, in order to make the whole model more robust, we treat the generated images of the target domain as an extension of the training data. We conduct extensive comparative experiments on multiple benchmark data sets. Experimental results verify that our method achieves improved results in high-quality face generation and LR face verification. In particular, our model DDAT reduces FID to 18.63 and 39.55 on the source and the target domain from 254.7 and 206.19 of the up-sampling results, respectively. Our method outperforms competing approaches by more than 10 percentage points in terms of face verification accuracy on multiple surveillance benchmarks. [ABSTRACT FROM AUTHOR]

• We propose a Behavior Regularized Prototypical Network (BR-ProtoNet) for few-shot image classification in semi-supervised scenarios. • BR-ProtoNet enables metric learning to benefit from readily-available unlabeled data. • We construct complementary constraints to regularize the model's behavior over the neighborhoods of training instances and along the interpolation paths among them. • The constructed regularization encourages the learnt embedding space to possess the property of proximity preservation. We propose a Behavior Regularized Prototypical Network (BR-ProtoNet) for few-shot image classification in semi-supervised scenarios. To learn a generalizable metric, we exploit readily-available unlabeled data and construct complementary constraints to regularize the model's behavior. Specifically, we match the label spaces between each episode and the whole training set. The predictions on the unlabeled data over different episodes can be aggregated to capture more reliable category information. We further construct new instances via adversarial perturbation and interpolation. These instances regularize the model's behavior over the neighborhoods of the original ones and along the interpolation paths among them. In addition, they ensure the learnt embedding space possesses the property of proximity preservation. The regularization of these aspects is incorporated into the optimization process of BR-ProtoNet on partially labeled data. We have conducted thorough experiments on multiple challenging benchmarks. The results suggest that the metric learning can significantly benefit from the proposed regularization, and thus leading to the state-of-the-art performance in semi-supervised few-shot image classification. [ABSTRACT FROM AUTHOR]

Unsupervised domain adaptation aims to address the problem in which the source data and target data are related but distributed differently. A widely-used two-stage strategy is to learn a domain-invariant subspace, and then train a cross-domain classifier on the resulting subspace. In this paper, we propose a single-stage domain adaption approach for joint subspace learning and discriminative learning. Specifically, a domain-invariant subspace and a cross-domain classifier are progressively learnt in a self-paced learning fashion. To avoid unlabeled target data dominating the overall loss and misleading model training, we progressively include more target data from "easy" to "complex" to optimize our model. Specifically, we propose an alternative optimization algorithm to efficiently find a reasonable solution for our task. Extensive experiments are conducted on multiple standard benchmarks to verify the effectiveness of the proposed approach. The results demonstrate that our model can outperform state-of-the-art non-deep domain adaptation methods. [ABSTRACT FROM AUTHOR]

• A simple yet effective unsupervised image translation method (SUIT)is proposed for domain adaptation on semantic segmentation, which avoids labor-intensive pixel-wise annotation. • The decoupled model design makes it moreflexible; the adaptation network be easily transplanted to other segmentation networks without repeating the adaptation process. • A model average skill is developed to improve the performance in domain adaptation context. The effectiveness of our proposed model is verified on multiple synthetic-to-real adaptation benchmarks. Image to image translation achieves superior performance with the advent of generative adversarial networks. In this paper, we propose a Simplified Unsupervised Image Translation (SUIT) model for domain adaptation on semantic segmentation. We adopt adversarial training for superior image generation, and design a novel semantic-content loss to enhance visual appearance preservation. Thus, the high-fidelity generated images with target-style can help the model generalize to the target domain. Besides, the semantic-content loss contains two components, which focus on label- and content-consistency, respectively. Both of them can be derived from existing modules of SUIT, which makes it simple yet suitable for domain adaptation on semantic segmentation tasks. Meanwhile, since the transformation network (generator) is decoupled from the segmentation network, the former can be easily transplanted to other semantic segmentation models. Extensive experimental results demonstrate that these translated images within SUIT can significantly improve performance of the model on the target domain, and our model with FCN8s-VGG16 architecture achieves around 13 percentage points improvement in terms of mIoU on multiple semantic segmentation adaptation benchmarks. [ABSTRACT FROM AUTHOR]