Search

Detecting cluttered and overlapping contraband items from baggage scans is one of the most challenging tasks, even for human experts. Recently, considerable literature has grown up around the theme of deep learning-based X-ray screening for localizing contraband data. However, the existing threat detection systems are still vulnerable to high occlusion, clutter, and concealment. Furthermore, they require exhaustive training routines on large-scale and well-annotated data in order to produce accurate results. To overcome the above-mentioned limitations, this paper presents a novel convolutional transformer system that recognizes different overlapping instances of prohibited objects in complex baggage X-ray scans via a distillation-driven incremental instance segmentation scheme. Furthermore, unlike its competitors, the proposed framework allows an incremental integration of new item instances while avoiding costly training routines. In addition to this, the proposed framework also outperforms state-of-the-art approaches by achieving a mean average precision score of 0.7896, 0.5974, and 0.7569 on publicly available GDXray, SIXray, and OPIXray datasets for detecting concealed and cluttered baggage threats. • This paper presents a novel incremental convolutional transformer model. • A β hyperparameter is introduced in the paper to control catastrophic forgetting. • A unique segmentation scheme is proposed to extract cluttered object instances. • The proposed system is thoroughly tested on three public X-ray datasets. [ABSTRACT FROM AUTHOR]

Machine learning in medical imaging often faces a fundamental dilemma, namely, the small sample size problem. Many recent studies suggest using multi-domain data pooled from different acquisition sites/centers to improve statistical power. However, medical images from different sites cannot be easily shared to build large datasets for model training due to privacy protection reasons. As a promising solution, federated learning, which enables collaborative training of machine learning models based on data from different sites without cross-site data sharing, has attracted considerable attention recently. In this paper, we conduct a comprehensive survey of the recent development of federated learning methods in medical image analysis. We have systematically gathered research papers on federated learning and its applications in medical image analysis published between 2017 and 2023. Our search and compilation were conducted using databases from IEEE Xplore, ACM Digital Library, Science Direct, Springer Link, Web of Science, Google Scholar, and PubMed. In this survey, we first introduce the background of federated learning for dealing with privacy protection and collaborative learning issues. We then present a comprehensive review of recent advances in federated learning methods for medical image analysis. Specifically, existing methods are categorized based on three critical aspects of a federated learning system, including client end, server end, and communication techniques. In each category, we summarize the existing federated learning methods according to specific research problems in medical image analysis and also provide insights into the motivations of different approaches. In addition, we provide a review of existing benchmark medical imaging datasets and software platforms for current federated learning research. We also conduct an experimental study to empirically evaluate typical federated learning methods for medical image analysis. This survey can help to better understand the current research status, challenges, and potential research opportunities in this promising research field. • Summarize existing methods from a system perspective. • Introduce different methods in a "question–answer" paradigm. • Introduce software platforms and benchmark datasets. • Conduct an experimental study. [ABSTRACT FROM AUTHOR]

The low-rank representation (LRR) method has attracted widespread attention due to its excellent performance in pattern recognition and machine learning. LRR-based variants have been proposed to solve the three existing problems in LRR: (1) the projection matrix is permanently fixed when dimensionality reduction techniques are adopted; (2) LRR fails to capture the local geometric structure; and (3) the solution deviates from the real low-rank solution. To address these problems, this paper proposes a low-rank representation with projection distance regularization (PDRLRR) via manifold optimization for clustering. In detail, we first introduce a low-dimensional projection matrix and a projection distance regularization term to fit the projected data automatically and capture the local structure of the data, respectively. Consequently, the projection matrix and representation matrix are obtained jointly. Then, we obtain a more accurate low-rank solution by minimizing the Schatten- p norm instead of the nuclear norm. Next, the projection matrix is optimized through a generalized Stiefel manifold. Extensive experiments demonstrate that our proposed method outperforms the state-of-the-art methods. • This paper proposes a novel PDRLRR model that can simultaneously address the three common problems in LRR. • Data dimensionality reduction technology is integrated into LRR, reducing the data dimensions while learning representation matrices. • For extracting the complete information, the projection distance regularization term is introduced to capture the global and local structure of the data. • The Schatten- p norm instead of the nuclear norm is employed to solve the rank minimization problem of the representation matrix, which can more accurately approximate the real low-rank solution. [ABSTRACT FROM AUTHOR]

Learning using privileged information (LUPI) has shown promise in improving supervised learning by embedding additional knowledge. However, its reliance on the assumption of readily available privileged information may not hold true in practical scenarios due to limitations in access or confidentiality. To address these challenges, this paper presents a novel weakly privileged learning (WPL) framework, integrating knowledge extraction methods within the LUPI context. An effective strategy is proposed to implement the WPL framework, where knowledge extraction techniques generate a weight matrix as weak privileged information. Extensive experiments employing various existing knowledge extraction techniques demonstrate that the proposed WPL outperforms traditional supervised learning and approaches the performance of standard privileged learning where privileged information is given in advance. This research establishes WPL as a promising learning paradigm, addressing limitations in privileged information availability and advancing the field of machine learning in practical settings. • Introduce a novel weakly privileged learning framework (WPL). • Knowledge extraction methods are used to generate weak privileged information. • Extensive experiments verify the superiority of the proposed methods. [ABSTRACT FROM AUTHOR]

Feature selection in classification is an important task in machine learning. Inspired by the success of Universum support vector machine proposed by Weston et al. on improving the classification ability of classical support vector machine, this paper considers a special type of Universum and further lets it play its role in both useful feature identification and separating hyperplane construction, aiming to improve both the feature selection ability and classification performance of Universum support vector machine. By introducing this special Universum, a redundant feature can be identified by observing whether some Universum sample is useful. In fact, we prove that by observing the dual solution of the optimization problem, useful features can be selected from a set satisfying some properties. Due to the introduction of these extra Universum samples, it needs to cope with a large-scale optimization problem. To improve the training efficiency, we modify the sequential minimal optimization algorithm and further combine it with the coordinate descent technique to solve the proposed model. Experimental results on artificial datasets, benchmark datasets, and text classification datasets demonstrate that the proposed method improves the classification performance of support vector machine and Universum support vector machine, and also has good feature selection ability. • Special Universum samples are introduced to fulfill both feature selection and classification. • It identifies a feature as useful if its corresponding Universum sample contributes to classification. • An effective algorithm by combining modified SMO and coordinate descent technique is designed. [ABSTRACT FROM AUTHOR]

Class incremental learning (CIL) has been proposed to solve the problem of learning to classify new classes while maintaining the performance on old classes. A typical strategy is to update the old classification model with entire new class data and a few old exemplars, which face serious performance degradation on old classes. The main reasons come down to class imbalance between old and new classes along with catastrophic forgetting towards old classes. Most existing CIL methods have proposed solving the above two issues in classification space, ignoring their adverse effects of overlapping between old and new classes and confusion among old classes in feature space. In this paper, we propose a rebalancing network with knowledge stability (RNKS), aiming to adequately retain the model performance on old classes in CIL by solving class imbalance and catastrophic forgetting in feature and classification space simultaneously. In detail, the proposed RNKS mainly consists of multi-proxies rebalancing (MPR) and hybrid knowledge distillation (HKD). MPR, focusing on class imbalance, employs multi-proxies metric learning to decrease the feature overlapping between old and new classes, together with balanced data sampling to correct the skewed decision boundary. HKD, coping with catastrophic forgetting, encourages the updated model to reproduce identical feature topologies and predictions of old classes as the old model via feature relation-based and response-based distillations. Experiments on CIFAR-100 and ILSVRC datasets demonstrate the effectiveness of this work against the state-of-the-art approaches. [ABSTRACT FROM AUTHOR]

Learning algorithms motivated by minimizing the sharpness of loss surface is a hot research topic in improving generalization. The existing methods usually solve a constrained min–max problem to minimize sharpness and find flat minima. However, most constraints (i.e., the neighborhood of the sharpness) are inappropriate, leading to sub-optimal results. This paper theoretically explores the optimal neighborhood from the view of Probably Approximately Correct-Bayesian (PAC-Bayesian) framework. A closed form of the optimal neighborhood is provided. This neighborhood is determined by the Hessian matrix and the scales of parameters. Then a generalization bound is derived that serves as a guiding principle in the design of the sharpness minimization algorithm. The Dynamic neighborhood-based Sharpness-Aware Minimization algorithm is proposed, which can adaptively adjust the neighborhood during the training process to gain better performance. Also, the algorithm is proved can convergent at the rate O (log T / T). Experimental results demonstrate that the proposed algorithm outperforms the other methods (e.g., accuracy ＋2.86% over baseline on CIFAR-100 for VGG-16). • A novel sharpness minimization algorithm is proposed to find flat minima. • The method can adaptively adjust the neighborhood of sharpness during the training. • Convergence analyses are deduced to guarantee the feasibility of the algorithm. • The algorithm outperforms other state-of-the-art baselines. • A new insight into the generalization theory and optimization algorithms is provided. [ABSTRACT FROM AUTHOR]

The adoption of artificial intelligence (AI) in medical imaging requires careful evaluation of machine-learning algorithms. We propose the use of a "deep virtual clinical trial" (DeepVCT) method to effectively evaluate the performance of AI algorithms. In this paper, DeepVCTs have been proposed to elucidate limitations of AI applications and predictions of clinical outcomes, avoiding biases in study designs. The DeepVCT method was used to evaluate the performance of nnU-Net models in assessing volumetric breast density (VBD) from digital breast tomosynthesis (DBT) images. In total, 2010 anatomical breast models were simulated. Projections were simulated using the acquisition geometry of a clinical DBT system. The projections were reconstructed using 0.1, 0.2, and 0.5 mm plane spacing. nnU-Net models were developed using the center-most planes of the reconstructions with the respective ground-truth. The results show that the accuracy of the nnU-Net improves significantly with DBT images reconstructed with 0.1 mm plane spacing (78. 4 × 205. 3 × 40. 1 mm3). The segmentations resulted in Dice values up to 0.84 with area under the receiver operating characteristic curve of 0.92. The optimization of plane spacing for VBD assessment was used as an exemplar of a DeepVCT application, allowing better interpretation of the input parameters and outcomes of the nnU-Net. Thus, DeepVCTs can provide evidence to predict the efficacy of AI algorithms using large-scale simulation-based data. [Display omitted] • DeepVCT provides simulation-based evidence to evaluate and optimize AI algorithms. • nnU-Net can be optimized to segment fibroglandular from adipose tissue. • Fraction of density affects volumetric breast density assessment. • Plane spacing also affects the volumetric breast density assessment. • Asymmetric evaluation metrics are necessary to handle imbalanced data. [ABSTRACT FROM AUTHOR]

Humans can easily distinguish the known and unknown categories and can recognize the unknown object by learning it once instead of repeating it many times without forgetting the learned object. Hence, we aim to make deep learning models simulate the way people learn. We refer to such a learning manner as O n L ine O pen W orld O bject D etection(OLOWOD). Existing OWOD approaches pay more attention to the identification of unknown categories, while the incremental learning part is also very important. Besides, some neuroscience research shows that specific noises allow the brain to form new connections and neural pathways which may improve learning speed and efficiency. In this paper, we take the dual-level information of old samples as perturbations on new samples to make the model good at learning new knowledge without forgetting the old knowledge. Therefore, we propose a simple plug-and-play method, called B rain-inspired S treaming D ual-level P erturbations(BSDP), to solve the OLOWOD problem. Specifically, (1) we first calculate the prototypes of previous categories and use the distance between samples and the prototypes as the sample selecting strategy to choose old samples for replay; (2) then take the prototypes as the streaming feature-level perturbations of new samples, so as to improve the plasticity of the model through revisiting the old knowledge; (3) and also use the distribution of the features of the old category samples to generate adversarial data in the form of streams as the data-level perturbations to enhance the robustness of the model to new categories. We empirically evaluate BSDP on PASCAL VOC and MS-COCO, and the excellent results demonstrate the promising performance of our proposed method and learning manner. • Designs an online open world object detection baseline. • Takes advantage of prototypes and statistics of old knowledge. • Proposes dual-level perturbations for capturing new and old knowledge. • Excellent results prove the promising performance of our method and learning manner. [ABSTRACT FROM AUTHOR]

Extensive research has been conducted on analyzing human movements, driven by its diverse practical applications such as human–robot interaction, human learning, and clinical diagnosis. However, the current state-of-the-art still encounters scientific challenges when it comes to modeling human movements. There are two key aspects that need to be addressed. Firstly, new models should consider the stochastic nature of human movement and the physical structure of the human body to accurately predict the patterns in full-body motion descriptors over time. Secondly, while deep learning algorithms have been utilized, they lack explainability in terms of predicting body posture sequences, making it essential to improve their comprehensible representation of human movement. This paper aims to tackle these challenges by presenting three innovative methods for creating explainable representations of human movement. The study formulates human body movement as a state-space model based on the Gesture Operational Model (GOM). Model parameters are estimated through either one-shot training employing Kalman Filters or data-intensive training utilizing artificial neural networks. The trained models are utilized for analyzing the dexterity of expert professionals in full-body movements, enabling the identification of dynamic associations between body joints and gesture recognition. Additionally, these models are employed to generate artificial professional movements. • Three innovative methods for explainable human movement models. • Body dexterity analysis: Models reveal how artisans and operators perform tasks. • Minimal motion descriptors to enhance task recognition. • Artificial generation of full-body professional movements. • Extensive experimentation with real-world datasets. [ABSTRACT FROM AUTHOR]

This paper addresses a machine learning problem often challenged by differences in the distributions of training and real-world data. We propose a framework that addresses the problem of underfitting in the ensembling method using pre-trained models and improves the performance and robustness of deep learning models through ensemble diversity. For the naive weight ensembling framework, we discovered that the ensembled models could not lie in the same loss basin under extreme domain shift conditions, suggesting that a loss barrier may exist. We used a fine-tuning step after the weighted ensemble to address the underfitting problem caused by the loss barrier and stabilize the batch normalization running parameters. We also inferred through qualitative analysis that the diversity of ensemble models affects domain generalization. We validate our method on a large-scale image dataset (ImageNet-1K) and chemical molecule data, which is suitable for testing with domain shift problems due to its data-splitting method. • We generalized the model via weight ensembling under an extreme distribution shift. • Our framework solves the underfitting problem of ensembles of pre-trained models. • We achieved up to 16.3% AUROC improvement over the baseline on the benchmark dataset. • Our model settles into a stable minima, which we verify on the loss surface. • We analyzed the correlation between model diversity and domain generalization. [ABSTRACT FROM AUTHOR]

Graph meta-learning has recently received significantly increased attention by virtue of its potential to extract common and transferable knowledge from learning different tasks on a graph. Existing methods for graph meta-learning usually leverage local subgraphs to transfer subgraph-specific information. However, they inherently face the challenge of imbalanced subgraphs due to inconsistent node density and different label distributions over local subgraphs. This paper proposes an adaptive graph meta-learning framework (AG-Meta) for learning the consistent and transferable representation of a graph in a way that can adapt to imbalanced subgraphs. Specifically, AG-Meta first learns the structural representation of subgraphs with various degrees using an Adaptive Graph Cascade Diffusion Network (AGCDN). AG-Meta then employs a prototype-consistency classifier to produce more accurate transferable inductive representations (also called prototypes) under few-shot settings with different label distributions of a subgraph. In the context of optimizing a model-agnostic meta-learner, a novel metric loss is finally introduced to achieve structural representation and prototype consistency. Extensive experiments are conducted to compare AG-Meta against baselines on five real-world networks, which validates that AG-Meta outperforms the state-of-the-art approaches. • A novel adaptive graph meta-learning framework is proposed. • An ensemble weight scheme is used to deal with inconsistency in densities and labels. • A prototypical-consistency loss is developed for graph meta-learning. • Our AG-Meta model outperforms several baselines. [ABSTRACT FROM AUTHOR]

Active learning alleviates labeling costs by selecting and labeling the most informative examples from an unlabeled pool. However, most existing active learning approaches estimate informativeness with uncalibrated confidence, resulting in unreliable informativeness estimation. These approaches generally ignored two significant issues caused by uncalibrated confidence methods. Firstly, the average uncalibrated confidence generated by modern neural networks is usually higher than the accuracy. Secondly, examples located near the decision boundaries are unstable during prediction when the target model updates parameters in the last several epochs, even throughout the training process. This phenomenon, caused by the forgetting characteristic of neural networks, has a significant impact on some specific models that estimate the informativeness by predicted probability vectors or pseudo labels. To address these issues, in this paper, we propose a novel active learning approach to reliably estimate informativeness with calibrated confidence. Specifically, we integrate the intermediate predictions for each unlabeled example, generated by the target model during the training process, to generate calibrated confidence. The calibrated confidence can capture a tendentious label from an indecisive subset of the class space. We show that the calibrated confidence with tendentiousness can maintain the ability of correct predictions. The empirical results demonstrate that our approach outperforms the state-of-the-art active learning methods on image classification tasks. • The active learning algorithm selects unstable examples by calibrated confidence.. • The confidence calibration method only requires training a single model. • Extensive experiments show the competitive performance of the proposed approach. [ABSTRACT FROM AUTHOR]

Lifelong learning, the ability to continually learn new concepts throughout our life, is a hallmark of human intelligence. Generally, humans learn a new concept by knowing what it looks like and what makes it different from the others , which are correlated. Those two ways can be characterized by generation and classification in machine learning respectively. In this paper, we carefully design a dynamically growing GAN called Introspective GAN (IntroGAN) that can perform incremental generation and classification simultaneously with the guidance of prototypes, inspired by their roles of efficient information organization in human visual learning and excellent performance in other fields like zero-shot/few-shot/incremental learning. Specifically, we incorporate prototype-based classification which is robust to feature change in incremental learning and GAN as a generative memory to alleviate forgetting into a unified end-to-end framework. A comprehensive benchmark on the joint incremental generation and classification task is proposed and our method demonstrates promising results. Additionally, we conduct comprehensive analyses over the properties of IntroGAN and verify that generation and classification can be mutually beneficial in incremental scenarios, which is an inspiring area to be further exploited. The code is available at https://github.com/TonyPod/IntroGAN. • Propose a benchmark to evaluate joint incremental generation and classification. • Propose an IntroGAN to address the joint incremental and classification task. • Introduce a fine-tuning technique that can accelerate incremental GAN training. • Superior performance on proposed benchmarks and more challenging scenarios. [ABSTRACT FROM AUTHOR]

Graph neural networks are increasingly becoming the framework of choice for graph-based machine learning. In this paper, we propose a new graph neural network architecture that substitutes classical message passing with an analysis of the local distribution of node features. To this end, we extract the distribution of features in the egonet for each local neighbourhood and compare them against a set of learned label distributions by taking the histogram intersection kernel. The similarity information is then propagated to other nodes in the network, effectively creating a message passing-like mechanism where the message is determined by the ensemble of the features. We perform an ablation study to evaluate the network's performance under different choices of its hyper-parameters. Finally, we test our model on standard graph classification and regression benchmarks, and we find that it outperforms widely used alternative approaches, including both graph kernels and graph neural networks. • We propose a novel GNN that analyzes local distributions of node features. • These are compared against learned ones using the histogram intersection kernel. • The similarity information is propagated to other nodes in the network. • This creates a message passing mechanism where the aggregation is non-linear. • Extensive experimental results demonstrate the competitiveness of our model. [ABSTRACT FROM AUTHOR]

Underwater images are usually covered with a blue–greenish colour cast, making them distorted, blurry or low in contrast. This phenomenon occurs due to the light attenuation given by the scattering and absorption in the water column. In this paper we present an image dewatering approach motivated upon the observation that the image formation model can be used to drive the learning process by constraining the loss function and making used of paired data. To this end, we employ a conditional generative adversarial network (cGAN) with two generators. Our Dual Generator Dewatering cGAN (DGD-cGAN) removes the haze and colour cast induced by the water column and restores the true colours of underwater scenes whereby the effects of various attenuation and scattering phenomena that occur in underwater images are tackled by the two generators. The first generator takes as input the underwater image and predicts the dewatered scene, while the second generator learns the underwater image formation process by implementing a custom loss function based upon the transmission and the veiling light components of the image formation model. Extensive experiments show that DGD-cGAN consistently delivers a margin of improvement as compared with state-of-the-art methods on several widely available datasets. • A dual generator conditional adversarial network performing underwater image restoration. • Constraining the colour restoration by the underwater image formation model. • Adversarial network training on a small dataset of real-world underwater images. • Implementing machine learning to cope with light attenuation and scattering. [ABSTRACT FROM AUTHOR]

Causal discovery reveals the true causal relationships behind data and discovering causal relationships from observed data is a particularly challenging problem, especially in large-scale datasets. The functional causal model is an effective method for causal discovery, but its time efficiency cannot be guaranteed. How to efficiently apply it to massive data still needs to be solved. In this paper, we propose a coreset construction for the additive noise model to accelerate causal discovery. According to the asymmetry characteristic of causality, samples were assigned different weights to construct the coreset. With the constructed coreset, we propose a Fast causal discovery algorithm based on the Additive Noise Model (FANM) to improve the time efficiency of the functional causal model while ensuring the result performance of causal discovery. Experiments on synthetic data and real-world data show that our proposed algorithm is much more time-efficient than the methods based on the functional causal model, and the runtime of FANM remains consistent as sample size increases while maintaining or exceeding the accuracy of the original nonlinear additive noise model. • New coresets proposed for the additive noise model greatly reduces the data size for causal discovery. • A time-efficient algorithm, FANM, is proposed for causal discovery based on the coresets. • The coreset construction is applied to causal graph learning algorithm, CDCSG, to show its effectiveness. • The proposed FANM is verified by both synthetic data and real-world data. [ABSTRACT FROM AUTHOR]

• Almost all these works above take the independent and identically distributed (i.i.d.) assumption, i.e., regarding all samples in various computer vision tasks as i.i.d., but they are not i.i.d. samples and the samples of many machine learning applications such as speech recognition, system diagnosis, and market forecasting are also proved to be non-i.i.d. samples. In particular, with the advent of the big data era, multi-class learning problems-such as text or image classification - can involve tens or hundreds of thousands of classes, and the samples are usually spatially-connected or temporally-correlated with their physically-connected neighbours, which can not be directly handled by these existing methods or algorithms. This implies that many samples are not satisfy the i.i.d. assumption or this assumption is very restrictive and can not be strictly justified in real-world problems. In addition, the AIO methods above are theoretically reasonable (i.e., Fisher consistent) and experiments indicate that they can lead to well-generalizing hypothesis, but efficient algorithms are lacking so far. Then three problems are posed: • Whether the AIO-MSVM algorithm based on non-i.i.d. samples is consistent? If this algorithm is consistent, what is the learning rate of this algorithm with non-i.i.d. samples? Whether the efficient AIO-MSVM algorithm can be proposed (or whether the AIO-MSVM algorithm with i.i.d. samples can be improved)? • The goal of this paper is to answer these problems above. Now, let' s highlight the main features of this paper. • The generalization bound of AIO-MSVM algorithm with u.e.M.c. samples is established and the fast learning rate is obtained. The AIO-MSVM algorithm with u.e.M.c. samples is proved to be consistent. To our knowledge, these studies are the first works on this topic. • A new algorithm, called AIO-MSVM algorithm based on q -times Markovian resampling (AIO-Mar- q), is proposed. The effectiveness and efficiency of the proposed algorithm are validated by experiments on public datasets. The generalization performance of "All-in-one" Multi-class SVM (AIO-MSVM) based on uniformly ergodic Markovian chain (u.e.M.c.) samples is considered. We establish the fast learning rate of AIO-MSVM algorithm with u.e.M.c. samples and prove that AIO-MSVM algorithm with u.e.M.c. samples is consistent. We also propose a novel AIO-MSVM algorithm based on q -times Markovian resampling (AIO-MSVM-MR), and show the numerical investigation on the learning performance of AIO-MSVM-MR based on public datasets. The experimental studies indicate that compared to the classical AIO-MSVM algorithm and other MSVM algorithms, the proposed AIO-MSVM-MR algorithm has not only smaller misclassification rate, but also less sampling and training total time. We present some discussions on the case of unbalanced training samples, the choices of q and two technical parameters, and present some explanations on the learning performance of the proposed algorithm. [ABSTRACT FROM AUTHOR]

Highlights • This paper proposed a frame work based on the inequality constrained optimization model to learn conditional probability table parameters by incorporating expert judgments and Dirichlet priors. • We further improve the proposed method by developing a constrained Bayesian Dirichlet prior. • Combined the proposed method, we provide an improved expectation maximum algorithm for learning conditional probability table parameters from incomplete data. • The contributed algorithm is tested on 13 well-known Bayesian networks, whose parameter number varies from 9 to 1157. The experiments show that the proposed method outperforms most of the existing parameter learning algorithms, especially when training data are extremely scarce. • A real facial action unit recognition case with incomplete data is conducted. The results show that the proposed method can build a more accurate Bayesian network for recognizing facial action units. Abstract Purely data-driven methods often fail to learn accurate conditional probability table (CPT) parameters of discrete Bayesian networks (BNs) when training data are scarce or incomplete. A practical and efficient means of overcoming this problem is to introduce qualitative parameter constraints derived from expert judgments. To exploit such knowledge, in this paper, we provide a constrained maximum a posteriori (CMAP) method to learn CPT parameters by incorporating convex constraints. To further improve the CMAP method, we present a type of constrained Bayesian Dirichlet priors that is compatible with the given constraints. Combined with the CMAP method, we propose an improved expectation maximum algorithm to process incomplete data. Experiments are conducted on learning standard BNs from complete and incomplete data. The results show that the proposed method outperforms existing methods, especially when data are extremely limited or incomplete. This finding suggests the potential effective application of CMAP to real-world problems. Moreover, a real facial action unit (AU) recognition case with incomplete data is conducted by applying different parameter learning methods. The results show that the recognition accuracy of respective recognition methods can be improved by the AU BN, which is trained by the proposed method. [ABSTRACT FROM AUTHOR]

Highlights • In this paper, we study the embedding of labels together with the group information with an objective to build an efficient multi-label classification. • We assume the existence of a low-dimensional space onto which the feature vectors and label vectors can be embedded. • We ensure that labels belonging to the same group share the same sparsity pattern in their low-rank representations. • The proposed method has three major stages namely (1) Identification of groups of labels; (2) Sparsity-invariant embedding of label groups; and (3) Embedding of feature matrix to the same low-rank space. • Extensive comparative studies validate the effectiveness of the proposed method against the state-of-the-art multi-label learning approaches. Abstract Multi-label learning is concerned with the classification of data with multiple class labels. This is in contrast to the traditional classification problem where every data instance has a single label. Due to the exponential size of output space, exploiting intrinsic information in feature and label spaces has been the major thrust of research in recent years and use of parametrization and embedding have been the prime focus. Researchers have studied several aspects of embedding which include label embedding, input embedding, dimensionality reduction and feature selection. These approaches differ from one another in their capability to capture other intrinsic properties such as label correlation, local invariance etc. We assume here that the input data form groups and as a result, the label matrix exhibits a sparsity pattern and hence the labels corresponding to objects in the same group have similar sparsity. In this paper, we study the embedding of labels together with the group information with an objective to build an efficient multi-label classifier. We assume the existence of a low-dimensional space onto which the feature vectors and label vectors can be embedded. In order to achieve this, we address three sub-problems namely; (1) Identification of groups of labels; (2) Embedding of label vectors to a low rank-space so that the sparsity characteristic of individual groups remains invariant; and (3) Determining a linear mapping that embeds the feature vectors onto the same set of points, as in stage 2, in the low-dimensional space. We compare our method with seven well-known algorithms on twelve benchmark data sets. Our experimental analysis manifests the superiority of our proposed method over state-of-art algorithms for multi-label learning. [ABSTRACT FROM AUTHOR]

Highlights • As far as we know, we are the first to apply CRC on domain adaptation problem. This paper proposed a jointly discriminative projection and dictionary learning method (JD2-CRC). The projection matrices and dictionary are learned according to the classification rule of CRC for improving the discriminability of collaborative representations. • Different to traditional optimization algorithm where a ratio trace problem is solved, we exploit the relationship of optimal variables, and optimize the trace ratio problem directly by a new trace ratio optimization method which is efficient and effective. • We conduct several experiments to verify the performances of the proposed algorithm. The experimental results show that the proposed optimization method is effective, and JD2-CRC achieves better performance than other state-of-the-art methods in domain adaptation problem. Abstract In recent years, collaborative representation-based classification (CRC) methods have shown impressive performance in many recognition tasks. However, when the training data have different distributions with the testing data, the performance of CRC will be degraded significantly. On the other hand, concatenating training data from different sources as a single data set will affect the performance of CRC, as the shift exists between the different source domains. To address these problems, in this paper, we propose a Jointly Discriminative projection and Dictionary learning for domain adaptive Collaborative Representation-based Classification method (JD2-CRC). As the distributions of different source domains may be dissimilar, the data from all domains are projected into a common feature subspace where the latent shared structures can be found. Then a compact dictionary is learned to represent the projected data well. To find the most suitable projection matrices and dictionary for CRC, we design the objective function of JD2-CRC,according to the classification rule of CRC in feature subspace, which minimizes the ratio of within-class reconstruction errors over between-class reconstruction errors. Different to traditional optimization methods, an effective optimization procedure is presented based on gradient descent. Thus, the obtained collaborative representations have a better discriminability and suit the classification rule of CRC well. The experimental results demonstrate that the proposed method can achieve superior performance against other state-of-the-art methods. [ABSTRACT FROM AUTHOR]

In extreme learning machines (ELM), the hidden node parameters are randomly generated and the output weights can be analytically computed. To overcome the bad feature extraction ability of the shallow architecture of ELM, the hierarchical ELM has been extensively studied as a deep architecture with multilayer neural network. However, the commonly used mean square error (MSE) criterion is very sensitive to outliers and impulsive noises, generally existing in real world data. In this paper, we investigate the correntropy to improve the robustness of the multilayer ELM and provide sparser representation. The correntropy, as a nonlinear measure of similarity, is robust to outliers and can approximate different norms (from ℓ 0 to ℓ 2 ). A new full correntropy based multilayer extreme learning machine (FC-MELM) algorithm is proposed to handle the classification of datasets which are corrupted by impulsive noises or outliers. The contributions of this paper are three-folds: (1) The MSE based reconstruction loss is replaced by the correntropy based loss function; In this way, the robustness of the ELM based multilayer algorithms is enhanced. (2) The traditional ℓ 1 -based sparsity penalty term is also replaced by a correntropy-based sparsity penalty term, which can further improve the performance of the proposed algorithm with a sparser representation of the data. The combination of (1) and (2) provides the correntropy-based ELM autoencoder. (3) The FC-MELM is proposed by using the correntropy-based ELM autoencoder as a building block. It is notable that the FC-MELM is trained in a forward manner, which means fine-tuning procedure is not required. Thus, the FC-MELM has great advantage in learning efficiently when compared with traditional deep learning algorithms. The good property of the proposed algorithm is confirmed by the experiments on well-known benchmark datasets, including the MNIST datasets, the NYU Object Recognition Benchmark dataset, and the Moore network traffic dataset. Finally, the proposed FC-MELM algorithm is applied to address Computer Aided Cancer Diagnosis. Experiments conducted on the well-known Wisconsin Breast Cancer Data (Diagnostic) dataset are presented and show that the proposed FC-MELM outperforms state-of-the-art methods in solving computer aided cancer diagnosis problems. [ABSTRACT FROM AUTHOR]

• The paper proposes the Presumably Correct Decision Sets for handling uncertainty. • Similarity neighborhoods define weak, borderline and strong members of each class. • Presumably correct and incorrect sets allow detecting difficult instances. • Numerical simulations illustrate the advantages of our model compared to rough sets. The paper presents the presumably correct decision sets as a tool to analyze uncertainty in the form of inconsistency in decision systems. As a first step, problem instances are gathered into three regions containing weak members, borderline members, and strong members. This is accomplished by using the membership degrees of instances to their neighborhoods while neglecting their actual labels. As a second step, we derive the presumably correct and incorrect sets by contrasting the decision classes determined by a neighborhood function with the actual decision classes. We extract these sets from either the regions containing strong members or the whole universe, which defines the strict and relaxed versions of our theoretical formalism. These sets allow isolating the instances difficult to handle by machine learning algorithms as they are responsible for inconsistent patterns. The simulations using synthetic and real-world datasets illustrate the advantages of our model compared to rough sets, which is deemed a solid state-of-the-art approach to cope with inconsistency. In particular, it is shown that we can increase the accuracy of selected classifiers up to 36% by weighting the presumably correct and incorrect instances during the training process. [ABSTRACT FROM AUTHOR]

• Cover all FSL papers from the 2000s to now, elaborate its evolution history. • Give an understandable hierarchical taxonomy for FSL approaches. • Categorize meta learning FSL methods and reveal their underlying relationship. • Analyze several emerging extensional research hotspots of FSL. • Conclude existing FSL applications and its future directions. The capability of learning and generalizing from very few samples successfully is a noticeable demarcation separating artificial intelligence and human intelligence. Despite the long history dated back to the early 2000s and the widespread attention in recent years with booming deep learning, few surveys for few sample learning (FSL) are available. We extensively study almost all papers of FSL spanning from the 2000s to now and provide a timely and comprehensive survey for FSL. In this survey, we review the evolution history and current progress on FSL, categorize FSL approaches into the generative model based and discriminative model based kinds in principle, and emphasize particularly on the meta learning based FSL approaches. We also summarize several recently emerging extensional topics of FSL and review their latest advances. Furthermore, we highlight the important FSL applications covering many research hotspots in computer vision, natural language processing, audio and speech, reinforcement learning and robotic, data analysis, etc. Finally, we conclude the survey with a discussion on promising trends in the hope of providing guidance and insights to follow-up researches. [ABSTRACT FROM AUTHOR]

Domain adaptation enables effective transfer between source and target domains with different distributions. The latest research focuses on open set domain adaptation; that is, the target domain contains unknown categories that do not exist in the source domain. The existing open set domain adaptation cannot realize the fine-grained recognition of unknown categories. In this paper, we propose an uncertainty analysis evidence model and design a distribution driven active transfer learning (DATL) algorithm. DATL realizes fine-grained recognition of unknown categories with no requirements on the source domain to contain the unknown categories. To explore unknown distributions, the uncertainty analysis evidence model was adopted to divide the high uncertainty space. To select critical instances, a cluster-diversity query strategy was proposed to identify new categories. To enrich the label categories of the source domain, a global dynamic alignment strategy was designed to avoid negative transfers. Comparative experiments with state-of-the-art methods on the standard Office-31/Office-Home/Office-Caltech10 benchmarks showed that the DATL algorithm: (1) outperformed its competitors; (2) realized accurate identification of unknown subcategories from a fine-grained perspective; and (3) achieved outstanding performance even with a very high degree of openness. • We propose a fine-grained open set domain adaptation issue. • We design a distribution driven active transfer learning algorithm. • We evaluate our method on benchmark datasets and demonstrate its effectiveness. [ABSTRACT FROM AUTHOR]

With the emerging of new data collection ways, the features are incremental and accumulated gradually. Due to the expansion of feature spaces, it is more common that there are unknown biases between the distribution of training and testing datasets. It is known as the unknown data selection bias, which belongs to the learning scenario with non-i.i.d samples. The performance of traditional approaches, which need the i.i.d. assumption, will be aggravated seriously. How to design an algorithm to address the problem of data selection bias in this feature incremental scenario is crucial but rarely studied. In this paper, we propose a feature incremental classification algorithm with causality. Firstly, we embed the confounding variable balance algorithm in causal learning into the prediction modeling and utilize the logical regression algorithm with balancing regular terms as a baseline. Then, to satisfy the special requirement of feature increment, we design a new regularizer, which maintains the consistency of the regression coefficients between the data in the current and previous stages. It retains the correlation between the old features and labels. Finally, we propose the Multiple Balancing Logistic Regression model (MBRLR) to jointly optimize the balancing regularizer and weighted logistic regression model with multiple feature sets. We also present theoretical results to show that our proposed algorithm can make precise and stable predictions. Besides, the numerical results also demonstrate that our MBRLR algorithm is superior to other methods. • We have pioneered the use of causal inference and proposed a novel method approach, i.e., MBRLR, for achieving stable and accurate classification in the feature increase scenario. • We have improved the sample reweighting method by redesigning the loss function and regularizer which could be formulated as a convex optimization problem and could be easily solved. • We have proved that the MBRLR algorithm can make a stable classification for feature incremental data in the non-i.i.d case. [ABSTRACT FROM AUTHOR]

• The first attempt to design a GEP-based codematrix generation algorithm. The individual structure of GEP allows the codematrix to be optimized in a much faster and more effective way. • The feature selection scheme is combined with the design of codematrix, to accomplish the feature selection and codematrix optimization simultaneously. • A greedy strategy-based adaptive optimization method to further improve the individual's quality to deal with hard classes by appending proper columns to the codematrix. The multi-class classification task is one of the most common tasks in machine learning. As a typical solution based on a partitioning strategy, Error-Correcting Output Codes (ECOC) can transform a multi-class classification problem into multiple binary classification problems. The key of ECOC is to construct an effective codematrix to represent a set of class decomposition schemes, which transforms a multiclass problem into a group of binary class problems. Consequently, the design of a fast and effective ECOC codematrix generation method is of great research significance and value for solving multi-class classification problems. In ECOC algorithms, the design of codematrix is treated as a combination problem between different code columns, in which the evolutionary algorithm shows a great advantage. Based on this consideration, the Gene Expression Programming (GEP) is applied to search for the codematrix with high performance because its expressive tree structure makes it well represent codematrcies for subsequent optimization operations. This paper proposes an adaptive ECOC algorithm based on Gene Expression Programming (GEP) and similarity measurement matrix, named GEP-ECOC. In our GEP, each individual represents a set of columns to form a random ECOC codematrix, which is optimized in the evolutionary process. Meanwhile, the crossover and mutation operations are modified to include a legality checking process to ensure that the generated codematrix satisfies the ECOC constraints. The GEP-based ECOC codematrix generation algorithm can quickly produce a codematrix with better performance, which ensures the efficiency of the algorithm to a certain extent. In addition, an adaptive algorithm based on a similarity measurement matrix is proposed to add new columns to the current codematrix, aiming to better handle hard classes. Our algorithm is compared with other algorithms on various data sets, and the experimental results confirm that our GEP-ECOC can balance the efficiency and performance of the algorithm and achieve higher performance. [ABSTRACT FROM AUTHOR]

Ensemble methods have shown to be more effective than monolithic classifiers, in particular when diversity holds among their components. How to enforce diversity in classifier ensembles has received much attention from machine learning researchers, yielding a variety of different techniques and algorithms. In this paper, a novel algorithm for ensemble classifiers is proposed, in which ensemble components are trained with focus on different regions of the sample space. In so doing, diversity is mainly a consequence of the intention to limit the scope of base classifiers. The algorithm proposed in this paper shares roots with several ensemble paradigms, in particular with random forests, as it generates forests of decision trees as well. As decision trees are trained with focus on specific subsets of the sample space, the resulting ensemble is in fact a forest of “local” trees. Comparative experimental results highlight that, on average, these ensembles perform better than other relevant kinds of ensemble classifiers, including random forests. [ABSTRACT FROM AUTHOR]

Multilayer network is a structure commonly used to describe and model the complex interaction between sets of entities/nodes. A three-layer example is the author-paper-word structure in which authors are linked by co-author relation, papers are linked by citation relation, and words are linked by semantic relation. Network embedding, which aims to project the nodes in the network into a relatively low-dimensional space for latent factor analysis, has recently emerged as an effective method for a variety of network-based tasks, such as collaborative filtering and link prediction. However, existing studies of network embedding both focus on the single-layer network and overlook the structural properties of the network, e.g., the degree distribution and communities, which are significant for node characterization, such as the preferences of users in a social network. In this paper, we propose four multilayer network embedding algorithms based on Nonnegative Matrix Factorization (NMF) with consideration given to four structural properties: whole network (NNMF), community (CNMF), degree distribution (DNMF), and max spanning tree (TNMF). Experiments on synthetic data show that the proposed algorithms are able to preserve the desired structural properties as designed. Experiments on real-world data show that multilayer network embedding improves the accuracy of document clustering and recommendation, and the four embedding algorithms corresponding to the four structural properties demonstrate the differences in performance on these two tasks. These results can be directly used in document clustering and recommendation systems. [ABSTRACT FROM AUTHOR]

In this paper we exploit the use of synthetic data for on-line handwritten gesture commands recognition with an emphasis on the problem of forgetting unused classes. For on-line learning, one of the most crucial moments of the processing is the initialization. In some applications the data is available and these can be fed to the learning model. However, in applications such as user-friendly handwritten gesture recognition, this scenario is not possible. Since from the user perspective it is better to let the user define his own symbols, the learning model is lacking in the amount of data at the initialization. Some strategies have been proposed to acquire synthetic handwritten gesture commands and use these for on-line learning. In this paper we exploit this technique further and focus on the forgetting of unused classes by applying a random buffer and Elastic Memory Learning (EML) to avoid this from happening. In the experiments we search for the proper amount of synthetic data produced for each sample as well as exploit the most appropriate time to stop the generation of synthetic data for learning purposes. We also investigate the influence of synthetic data on forgetting when using the proposed EML. We base the generation of synthetic gesture commands on Kinematic Theory. [ABSTRACT FROM AUTHOR]

In this paper, we address a machine learning task using encrypted training data. Our basic scenario has three parties: Data Owners, who own private data; an Application, which wants to train and use an arbitrary machine learning model on the Users’ data; and an Authorization Server, which provides Data Owners with public and secret keys of a partial homomorphic cryptosystem (that protects the privacy of their data), authorizes the Application to get access to the encrypted data, and assists it in those computations not supported by the partial homomorphism. As machine learning model, we have selected the Support Vector Machine (SVM) due to its excellent performance in supervised classification tasks. We evaluate two well known SVM algorithms, and we also propose a new semiparametric SVM scheme better suited for the privacy-protected scenario. At the end of the paper, a performance analysis regarding the accuracy and the complexity of the developed algorithms and protocols is presented. [ABSTRACT FROM AUTHOR]

• We propose a novel Self-taught Multi-view Spectral Clustering (SMSC) framework, which can directly obtain a discrete indicator matrix. • Our SMSC is devised to integrate adaptive weighting, Laplacian embedding (spectral embedding), consensus graph learning, and discrete indicator matrix learning into a unified framework. • Our SMSC can dynamically update the consensus graph based on the feedback of the discrete cluster label matrix and vice versa. • An efficient algorithm is proposed to solve the challenging discrete optimization problem alternately. By integrating multiple views, i.e., multi-view learning (ML), we can discover the underlying data structures so that the performance of learning tasks can improve. As a basic and important branch of ML, multi-view clustering has achieved great success recently in pattern recognition and machine learning communities. Most existing multi-view spectral clustering methods heavily adopt the relax-and-discretize strategy to obtain discrete cluster labels (clustering results), i.e., using predefined similarity graphs to learn a consensus Laplacian embedding shared by all views for K -means clustering. However, the above clustering strategy may significantly affect clustering performance since there is information loss between independent steps. In this paper, we establish a novel Self-taught Multi-view Spectral Clustering (SMSC) framework to address the above issue. As the main contributions of this paper, we provide two versions of SMSC based on convex combination and centroid graph fusion schemes. Specifically, a self-taught mechanism is introduced in SMSC, which can effectively feedback the manifold structure induced by Laplacian embedding and the cluster information hidden in the discrete indicator matrix to learn an optimal consensus similarity graph for graph partitioning. The effectiveness of the proposed methods has been evaluated on real-world multi-view datasets, and experimental results show that our methods outperform other state-of-the-art baselines. [ABSTRACT FROM AUTHOR]

• The proposed dual-branch network separates knowledge to avoid gradient conflicts. • The proposed adaptive feature fusion module aggregates knowledge from two branches. • The proposed model compression stage preserves the dual-branch structure. • Experiments show that our method outperforms other methods on different benchmarks. Most existing class incremental learning methods rely on storing old exemplars to avoid catastrophic forgetting. However, these methods inevitably face the gradient conflict problem, the inherent conflict between new streaming knowledge and existing knowledge in the gradient direction. To alleviate gradient conflict, this paper reuses the previous knowledge and expands the branch to accommodate new concepts instead of fine-tuning the original models. Specifically, this paper designs a novel dual-branch network called Knowledge Aggregation Networks. The previously trained model is frozen as a branch to retain existing knowledge, and a consistent trainable network is constructed as the other branch to learn new concepts. An adaptive feature fusion module is adopted to dynamically balance the two branches' information during training. Moreover, a model compression stage maintains the dual-branch structure. Extensive experiments on CIFAR-100, ImageNet-Sub, and ImageNet show that our method significantly outperforms the other methods and effectively balances stability and plasticity. [ABSTRACT FROM AUTHOR]

• A constrained sparse graph construction method is proposed. • The method does not require a predefined affinity matrix. • The proposed constraints impose edge weights smoothness. • The proposed constraints lead to a structured sparse graph. • Performance is assessed on graph-based label propagation and embedding. Sparse representation is a useful tool in machine learning and pattern recognition area. Sparse graphs (graphs constructed using sparse representation of data) proved to be very informative graphs for many learning tasks such as label propagation, embedding, and clustering. It has been shown that constructing an informative graph is one of the most important steps since it significantly affects the final performance of the post graph-based learning algorithm. In this paper, we introduce a new sparse graph construction method that integrates manifold constraints on the unknown sparse codes as a graph regularizer. These constraints seem to be a natural regularizer that was discarded in existing state-of-the art graph construction methods. This regularizer imposes constraints on the graph coefficients in the same way a locality preserving constraint imposes on data projection in non-linear manifold learning. The proposed method is termed Sparse Graph with Laplacian Smoothness (SGLS). We also propose a kernelized version of the SGLS method. A series of experimental results on several public image datasets show that the proposed methods can out-perform many state-of-the-art methods for the tasks of label propagation, nonlinear and linear embedding. [ABSTRACT FROM AUTHOR]

• We develop a semi-supervised learning framework of hierarchically structured networks. • The proposed method utilizes matrix sparseness and approximations to solve issues on computational complexity, sparseness, and scalability arising from hierarchically structured networks. • We provide analyses on error bounds and complexity to show suitability of the proposed method on semi-supervised learning framework. • The experimental results show that the proposed algorithms perform well with hierarchically structured data, and, outperform an ordinary semi-supervised learning algorithm. A set of data can be obtained from different hierarchical levels in diverse domains, such as multi-levels of genome data in omics, domestic/global indicators in finance, ancestors/descendants in phylogenetics, genealogy, and sociology. Such layered structures are often represented as a hierarchical network. If a set of different data is arranged in such a way, then one can naturally devise a network-based learning algorithm so that information in one layer can be propagated to other layers through interlayer connections. Incorporating individual networks in layers can be considered as an integration in a serial/vertical manner in contrast with parallel integration for multiple independent networks. The hierarchical integration induces several problems on computational complexity, sparseness, and scalability because of a huge-sized matrix. In this paper, we propose two versions of an algorithm, based on semi-supervised learning, for a hierarchically structured network. The naïve version utilizes existing method for matrix sparseness to solve label propagation problems. In its approximate version, the loss in accuracy versus the gain in complexity is exploited by providing analyses on error bounds and complexity. The experimental results show that the proposed algorithms perform well with hierarchically structured data, and, outperform an ordinary semi-supervised learning algorithm. [ABSTRACT FROM AUTHOR]

In this paper, we develop an approach to exploit kernel methods with data lying on the m -D Kendall shape space. When data arise in a finite-dimensional curved Riemannian manifold, as in this case, the usual Euclidean computer vision and machine learning algorithms must be treated carefully. A good approach is to use positive definite kernels on manifolds to embed the manifold with its corresponding metric in a high-dimensional reproducing kernel Hilbert space, where it is possible to utilize algorithms developed for linear spaces. Different Gaussian kernels can be found in the literature on the 2-D Kendall shape space to perform this embedding. The main novelty of this work is to provide a Gaussian kernel for the m -D Kendall shape space. This new Kernel coincides in the case m = 2 with the Gaussian kernels most widely used in the Kendall planar shape space and allows to define an embedding of the m -D Kendall shape space into a reproducible kernel Hilbert space. As far as we know, the complexity of the m -D Kendall shape space has meant that this embedding has not been addressed in the literature until now. This methodology will be tested on a machine learning problem with a simulated and a real data set. • Proposed kernels in the m-dimensional Kendalls space. • Shapes embedded in a high-dimensional Reproducing Kernel Hilbert Space. • Unification and generalization of Gaussian kernels defined in the space of planar shapes. • Supervised classification problems using a support vector machine algorithm. [ABSTRACT FROM AUTHOR]

Medical data sharing across institutes is crucial to large-scale multi-center studies and the development of real-world AI applications but suffers from serious privacy issues. A promising solution to address this challenge is federated learning, which typically aggregates a global model from heterogeneous data spread across numerous clients without exchanging data. However, the traditional federated learning algorithm (i.e. , FedAvg) merely aggregates the locally distributed models according to the amount of data on each client and lacks the consideration of data heterogeneity. In this paper, we propose a novel Federated Adaptive Reweighting (FedAR) algorithm for medical image classification. FedAR employs a flexible re-weighting scheme that can balance adaptively the contributions of the amount of data and the performance of the local model on each client to the weight of that client. Specifically, we allow the amount of local data to contribute more to the weight of each client in the early training stage and let the performance of the local model play a more important role in the late stage. We have evaluated the proposed FedAR algorithm against the locally trained model, globally trained baseline, and two existing federated learning algorithms on the ISIC2018 dataset and Chest X-ray14 dataset under the settings with a variable number of clients. Our results suggest that FedAR is an effective federated learning algorithm that substantially outperforms existing federated learning approaches. • Propose FedAR, based on federated learning, for medical image classification. • Develop an adaptive reweighting scheme to address the heterogeneity issue. • Demonstrate the effectiveness of FedAR on the ISIC2018 and Chest X-ray14 datasets. • FedAR improves the federated learning with limited extra communication costs. [ABSTRACT FROM AUTHOR]

• The study uses cutting-edge deep learning algorithms for recognizing facial asymmetry, a key stroke symptom. • Our method significantly improves existing stroke detection techniques by providing automated, real-time assessment of potential stroke victims. • The research incorporates the latest findings in the field of pattern recognition, ensuring that our system is current and relevant. • We present a unique system that does not require any specialized equipment for detection, potentially improving the accessibility of stroke detection. • The paper demonstrates how this automated detection system could be a crucial first step in stroke segmentation, thus contributing to the early intervention and treatment of stroke. Strong efforts have been undertaken to enhance the diagnosis and identification of diseases that cause facial paralysis, such as Bell's palsy and stroke, because of their detrimental social effects. Stroke is one of the most serious and potentially fatal conditions among the major cardiovascular disorders. We are introducing a deep-learning-based method for early diagnosis of facial paralysis diseases such as stroke and Bell's palsy. Recognizing the costs associated with traditional diagnostic techniques like magnetic resonance tomography (MRI) and computed tomography (CT) scan images, our model employs a multi-task network, integrating face parsing, facial asymmetry parsing, and category enhancement. Spatial inconsistencies are addressed via a depth-map estimation module that leverages an instance-specific kernel approach. To clarify the boundaries of facial components, we use category edge detection with a foreground attention module, generating generic geometric structures and detailed semantic cues. Our model is trained on two datasets, comprising individuals with regular smiles and those with one-sided facial weakness. This cost-effective, easily accessible solution can streamline the diagnostic process, minimizing data gaps, and reducing needless rescreening and intervention costs. [ABSTRACT FROM AUTHOR]

Online portfolio selection (OPS) has received increasing attention from machine learning and quantitative finance communities. Despite their effectiveness, the pioneering OPS methods have several key limitations. First, price predictions are usually based on predetermined trends, which is inadequate for a fast-changing market patterns. Second, each asset is treated individually, ignoring the pervading relevance among the assets. Third, the risk terms are usually missing or inappropriate in optimizations. This paper proposes a novel OPS method, namely, the online low-dimension ensemble method, to overcome the limitations. Motivated by the stylized facts for the co-movements of assets, the financial market is regarded as a high-dimensional dynamical system (HDS), and a large number of low-dimensional subsystems (LDSs) are randomly generated from the HDS to extract the correlation information among the assets. The assets' price predictions are first made using these LDSs and then aggregated to formulate the final prediction using ensemble learning techniques. Thanks to the particular merits brought by our predicting scheme, we also develop a novel high-dimensional covariance matrix estimation/prediction method for short-term data, efficiently assessing the instantaneous risk of the projected portfolios. Compared with state-of-the-art methods, our approach obtains more accurate predictions as the correlation information is fully exploited. With the predictive instantaneous risk assessment, a more appropriate optimization problem is proposed, substantially improving the OPS setting and leading to significantly better investment performance. Therefore, this study develops a flexible and promising approach to learning fast-changing market patterns and demonstrates that the high-dimensional feature of the market is a crucial information source for financial modeling with short-term data rather than a barrier in the conventional sense. Extensive experiments on real-world datasets are conducted to illustrate our method further. • Proposing a novel approach to learn short-term high-dimensional pattern. • Developing a novel high-dimensional covariance matrix prediction/estimation method. • Substantially improving the optimization settings of OPS. [ABSTRACT FROM AUTHOR]

• We propose a novel concept of Paired Contrastive Feature (PCF) for highly reliable writer-independent signature verification by converting a task with pairwise contrastive evaluation into a task with sample-wise evaluation. • Two highly reliable machine-learning frameworks, top-rank learning and learning with rejection, are applied to the writer-independent signature verification task for the first time by constructing PCF to the authors' best knowledge. • We validated the reliability and effectiveness of the proposed PCF with highly reliable methods through experiments with multiple evaluation metrics. [Display omitted] Signature verification requires high reliability. Especially in the writer-independent scenario with the skilled-forgery-only condition, achieving high reliability is challenging but very important. In this paper, we propose to apply two machine learning frameworks, learning with rejection and top-rank learning, to this task because they can suppress ambiguous results and thus give only reliable verification results. Since those frameworks accept a single input, we transform a pair of genuine and query signatures into a single feature vector, called Paired Contrastive Feature (PCF). PCF internally represents similarity (or discrepancy) between the two paired signatures; thus, reliable machine learning frameworks can make reliable decisions using PCF. Through experiments on three public signature datasets in the offline skilled-forgery-only writer-independent scenario, we evaluate and validate the effectiveness and reliability of the proposed models by comparing their performance with a state-of-the-art model. [ABSTRACT FROM AUTHOR]

• A fast and privacy-preserving training algorithm for deep autoencoders is proposed. • The encoder block of the autoencoder is trained using Singular Value Decomposition. • The decoder block of the autoencoder is trained layer by layer. • The non-iterative nature of the proposed training algorithm makes it faster than other approaches. • The method is especially suitable for federated learning and edge computing environments. This paper presents a novel, fast and privacy preserving implementation of deep autoencoders. DAEF (Deep AutoEncoder for Federated learning), unlike traditional neural networks, trains a deep autoencoder network in a non-iterative way, which drastically reduces training time. Training can be performed incrementally, in parallel and distributed and, thanks to its mathematical formulation, the information to be exchanged does not endanger the privacy of the training data. The method has been evaluated and compared with other state-of-the-art autoencoders, showing interesting results in terms of accuracy, speed and use of available resources. This makes DAEF a valid method for edge computing and federated learning, in addition to other classic machine learning scenarios. [ABSTRACT FROM AUTHOR]

• We propose a federated contrastive learning method, which uses the feature representation learned by the global model and the local model for contrastive learning, so that the final model can gradually approach the global model in the training process to obtain better model generalization. • In view of the current problem of data heterogeneity in real medical scenes, our method effectively reduces its impact on experimental results, that is, it can show high performance and generalization in the scenario of data heterogeneity. • We conducted experiments on two medical image classification datasets, i.e., skin lesion classification dataset for melanoma and chest X-ray dataset for detecting COVID-19. Extensive experiments demonstrate the efficacy of our approach. Federated learning, which allows distributed medical institutions to train a shared deep learning model with privacy protection, has become increasingly popular recently. However, in practical application, due to data heterogeneity between different hospitals, the performance of the model will be degraded in the training process. In this paper, we propose a federated contrastive learning (FedCL) approach. FedCL integrates the idea of contrastive learning into the federated learning framework. Specifically, it combines the local model and the global model for contrastive learning, so that the local model gradually approaches the global model with the increase of communication rounds, which improves the generalization ability of the model. We validate our method on two public datasets. Extensive experiments show that our method is superior to other federated learning algorithms in medical image classification. [ABSTRACT FROM AUTHOR]

• We propose a general-purpose nonlinear dimensionality reduction method named Vec2vec by extending the representation learning of words to numeric matrices, which has lower computational complexity than most existing dimensionality reduction methods on high-dimensional data. • We design a tailored objective function to preserve the pairwise similarities of data points and define the contexts of data points as random walk sequences in the neighborhood similarity graph, which serve as samples for model training. • We propose a more lightweight adaption of Vec2vec , which is referred to as AVec2vec. AVec2vec further reduces the computational complexity of Vec2vec with marginal loss in structural information. • We conduct extensive experiments for data classification and clustering on nine popular real-world image and text datasets to evaluate the performance of our method. The results show that Vec2vec outperforms several state-of-the-art dimensionality reduction methods in statistical hypothesis tests, and is more efficient than UMAP on high-dimensional data. The results also show that AVec2vec needs less computational time than Vec2vec with little performance loss in data classification and clustering tasks. Dimensionality reduction has many applications in data visualization and machine learning. Existing methods can be classified into global ones and local ones. The global methods usually learn the linear relationship in data, while the local ones learn the manifold intrinsic geometry structure, which has a significant impact on pattern recognition. However, most of existing local methods obtain an embedding with eigenvalue or singular value decomposition, where the computational complexities are very high in a large amount of high-dimensional data. In this paper, we propose a local nonlinear dimensionality reduction method named Vec2vec , which employs a neural network with only one hidden layer to reduce the computational complexity. We first build a neighborhood similarity graph from the input matrix, and then define the context of data points with the random walk properties in the graph. Finally, we train the neural network with the context of data points to learn the embedding of the matrix. We conduct extensive experiments of data classification and clustering on nine image and text datasets to evaluate the performance of our method. Experimental results show that Vec2vec is better than several state-of-the-art dimensionality reduction methods, except that it is equivalent to UMAP on data clustering tasks in the statistical hypothesis tests, but Vec2vec needs less computational time than UMAP in high-dimensional data. Furthermore, we propose a more lightweight method named Approximate Vec2vec (AVec2vec) with little performance degradation, which employs an approximate method to build the neighborhood similarity graph. AVec2vec is still better than some state-of-the-art local dimensionality reduction methods and competitive with UMAP on data classification and clustering tasks in the statistical hypothesis tests. [ABSTRACT FROM AUTHOR]

• We developed two distinct strategies and two algorithms that address certain challenges associated with employing adversarial perturbation algorithms in the context of MIL. Furthermore, the efficacy of the algorithm has been verified across diverse input data scenarios. • The resulting adversarial bags and the proposed methods can be deemed valuable knowledge or prior experience for future learners. This is because MI-CAP can generate perturbations that remain unaffected by other bags, and the perturbations generated by MI-UAP can be conveniently stored as part of the knowledge base. • We have demonstrated the generalizability of these perturbations through experiments, illustrating that perturbations generated by one neural network can fool other networks as well. Furthermore, we have put forth a straightforward strategy to mitigate the impact of these perturbations as much as possible. Multi-instance learning (MIL) is a recent machine learning paradigm which is immensely useful in various real-life applications, like image analysis, video anomaly detection, text classification, etc. It is well known that most of the existing machine learning classifiers are highly vulnerable to adversarial perturbations. Since MIL is a weakly supervised learning, where information is available for a set of instances, called bag and not for every instance, adversarial perturbations can be fatal. In this paper, we have proposed two adversarial perturbation methods to analyze the effect of adversarial perturbations to interpret the vulnerabilities of MIL methods. Out of the two algorithms, one can be customized for every bag, and the other is a universal one, which can affect all bags in a given data set and thus has some generalizability. Furthermore, through simulations, we have demonstrated the efficacy of the proposed algorithms in fooling state-of-the-art MIL approaches, such that these models make incorrect predictions regarding the label assigned to the bag. Finally, we have discussed, through experiments, about taking care of these kind of adversarial perturbations through a simple strategy. Source codes are available at https://github.com/InkiInki/MI-UAP. [ABSTRACT FROM AUTHOR]

• Using ℓ 20 norm constraint to obtain a pure sparse subspace. • Providing an algorithm to obtain a closed-form solution for ℓ 20 norm constraint problem. • The high efficiency of algorithm is guaranteed by the convergence proof. • Exploiting local discriminative structure to learn a better subspace. Clustering is a fashion method applied in machine learning tasks. However, high dimensional data brings many obstacles for clustering approaches. To address such a problem, the unsupervised feature selection (UFS) method can be incorporated into clustering to reduce dimensionality. In general, most of the UFS methods adopt ℓ 2 , 1 -norm for subspace sparsity learning. However, its sparsity highly relies on the setting of trade-off parameter, which may lead to instability of ranking results and the difficulty in obtaining the optimal solution of projection matrix. In this paper, we propose to directly learn an absolutely row-sparsity subspace via the ℓ 2 , 0 -norm constraint, called Sparse constraint and Local learning for Unsupervised Feature Selection (SLUFS). It is an ideal sparse subspace constraint which can overcome the drawbacks of the ℓ 2 , 1 -norm. However, optimizing the ℓ 2 , 0 -norm constraint is an NP-hard problem, and at present, only some approximate solutions can be given, but the convergence can not be guaranteed. To tackle this challenge, we design a novel alternative iterative algorithm to directly optimize the ℓ 2 , 0 -norm based model. Most importantly, our strategy can obtain a closed-form solution with strict convergence guarantee. Comprehensive experiments are conducted on several real-world datasets to evaluate the performance of SLUFS with comparison to several related state-of-the-art methods. [ABSTRACT FROM AUTHOR]

• We propose a tree representation based on constrained matrices that leads to two new generic linear programming formulations of the optimal decision tree problem allowing to easily integrate domain-knowledge constraints. • Both formulations give the possibility for users to easily control complexity through the constraints on the number of leaf nodes and the maximum depth of the decision tree. • We theoretically show that elements of these matrices give the ability to easily formalise domain knowledge as constraints to improve trustworthiness. When applied to critical domains, machine learning models usually need to comply with prior knowledge and domain-specific requirements. For example, one may require that a learned decision tree model should be of limited size and fair, so as to be easily interpretable, trusted, and adopted. However, most state-of-the-art models, even on decision trees, only aim to maximising expected accuracy. In this paper, we propose a framework in which a diverse family of prior and domain knowledge can be formalised and imposed as constraints on decision trees. This framework is built upon a newly introduced tree representation that leads to two generic linear programming formulations of the optimal decision tree problem. The first one targets binary features, while the second one handles continuous features without the need for discretisation. We theoretically show how a diverse family of constraints can be formalised in our framework. We validate the framework with constraints on several applications and perform extensive experiments, demonstrating empirical evidence of comparable performance w.r.t. state-of-the-art tree learners. [ABSTRACT FROM AUTHOR]

A new edge detection technique using transformation groups based G-lets filters is proposed in this paper. Discretizing gradients seem to produce discontinuity in classic edge detectors. No particular filter is capable of identifying meaningful edges at all scales and it increases computations with a multiscale approach. It is a challenge to get localized edges without spurious ones due to noise and integrate the obtained edges into meaningful object boundaries. Without breaking edge continuity and strictly localizing edges requires that filters do not blur the image during preprocessing. G-lets filters are found to be capable of performing well in most type of images including natural, noisy, low resolution and synthetic. In this paper, an edge detection algorithm using G-lets filters which are built by direct factorization of linear transformation matrices using irreducible representations is proposed. A multiresolution approach is shown to enhance the possibility of detecting faint edges. An edge tracing algorithm is presented to produce the edge image. The computational cost involved is comparatively lesser than existing filters. It is found that the geometries in the original image are preserved in the edge image. The edge tracing algorithm is capable of constructing object boundaries without the inner textures in a way that is not completely dependent on intensity thresholding. G-lets filters and the edge operator is found to be a promising algorithm for drastically bringing down the computations needed for realtime applications. The results are compared with BSDS500 boundary detection dataset using pb and global pb detectors. [ABSTRACT FROM AUTHOR]

Image registration is one of the major technologies of the medical image processing and analysis. According to the nature of the transformation, image registration can be categorized into two main classes: Rigid Registration and Non-rigid Registration. In this paper, a novel learning-based dissimilarity metric and its approximation are proposed and respectively adopted to the rigid and non-rigid registration of medical images. This novel metric utilizes Bhattacharyya Distances to measure the dissimilarity of the testing image pairs by incorporating the expected intensity distributions which is learnt from the registered training image pairs. In order to test the robustness and accuracy of the proposed metric in the multi-modal rigid registration, one thousand eight hundred randomized CT–T1 registrations were performed and evaluated by the Retrospective Image Registration Evaluation (RIRE) project. For the non-rigid registration, the approximation of the proposed dissimilarity metric is adopted with the Markov Random Field (MRF) modeled non-rigid registration approach. Non-rigid registration experiments of 2D synthetic images with artificial deformation fields and 3D images from the Simulated Brain Database were performed and evaluated respectively in this paper. The experimental results of both rigid and non-rigid medical image registrations demonstrated that the proposed dissimilarity metric outperforms the state-of-the-art approach (Mutual Information) and the close related approach (Kullback–Leibler Distances). [ABSTRACT FROM AUTHOR]

Multi-label classification problems usually occur in tasks related to information retrieval, like text and image annotation, and are receiving increasing attention from the machine learning and pattern recognition fields. One of the main issues under investigation is the development of classification algorithms capable of maximizing specific accuracy measures based on precision and recall. We focus on the widely used F measure, defined for binary, single-label problems as the weighted harmonic mean of precision and recall, and later extended to multi-label problems in three ways: macro-averaged, micro-averaged and instance-wise. In this paper we give a comprehensive survey of theoretical results and algorithms aimed at maximizing F measures. We subdivide it according to the two main existing approaches: empirical utility maximization, and decision-theoretic. Under the former approach, we also derive the optimal (Bayes) classifier at the population level for the instance-wise and micro-averaged F , extending recent results about the single-label F . In a companion paper we shall focus on the micro-averaged F measure, for which relatively fewer solutions exist, and shall develop novel maximization algorithms under both approaches. [ABSTRACT FROM AUTHOR]

• This paper proposes a method to replace the linear mapping-based spectral analysis with neural networks, which is more appropriate to learn the nonlinear mapping relationship between the original data and the pseudo cluster label. • Through constructing two different adaptive graph constraints, the mutual interaction between the NNs-based pseudo label learning module and the self-expression module is facilitated to explore the feature selection priority for the raw data. • Through a lot of well-designed experiments, the effectiveness of the method proposed in this paper has been fully proven. Unsupervised feature selection (UFS), which selects the most important feature subset and eliminates the unnecessary information for the upcoming data analysis, is a significant problem in machine learning and has been explored for years. Most UFS methods map features into a pseudo label space by multiplying a projection matrix constrained with sparsity to learn the mapping from the features to the labels. However, the mapping relationship is usually not linear, and linear regression may result in a suboptimal selection. To address this issue, we propose a novel UFS method, called neural networks embedded self-expression (NNSE). NNSE replaces the linear regression of traditional spectral analysis methods with neural networks to learn the pseudo label space. Besides, we embed neural networks into the self-expression model to improve the representative ability by preserving the local structure with an adaptive graph regularization module. Then we propose an efficient alternative iterative algorithm to solve the proposed model. Experimental results on 8 public datasets show NNSE outperforms the other state-of-the-art methods. Moreover, experimental results are also presented to show the convergence of the proposed method. The source code is available at: https://github.com/misteru/NNSE. [ABSTRACT FROM AUTHOR]