Search

Automatic identity recognition from ear images is an active research topic in the biometric community. The ability to secretly acquire images of the ear remotely and the stability of the ear shape over time make this technology a promising alternative for surveillance, authentication, and forensic applications. In recent years, significant research has been conducted in this area. Nevertheless, challenges remain that limit the commercial use of this technology. Several phases of the ear recognition system have been studied in the literature, from ear detection, normalization, and feature extraction to classification. This paper reviews the most recent methods used to describe and classify biometric features of the ear. We propose a first taxonomy to group existing approaches to ear recognition, including 2D, 3D, and combined 2D and 3D methods, as well as an overview of historical advances in this field. It is well known that data and algorithms are the essential components in biometrics, particularly in-ear recognition. However, early ear recognition datasets were very limited and collected in laboratory with controlled environments. With the wider use of deep neural networks, a considerable amount of training data has become necessary if acceptable ear recognition performance is to be achieved. As a consequence, current ear recognition datasets have increased significantly in size. This paper gives an overview of the chronological evolution of ear recognition datasets and compares the performance of conventional vs. deep learning methods on several datasets. We proposed a second taxonomy to classify the existing databases, including 2D, 3D, and video ear datasets. Finally, some open challenges and trends are debated for future research. [ABSTRACT FROM AUTHOR]

Text classification is an essential and classic problem in natural language processing. In recent years, Graph Convolutional Networks (GCNs) have been widely applied to text classification tasks. However, there are still three critical challenges in practical applications: (1) Limited explanatory ability for domain-specific terminology; (2) Imbalanced sample distribution leading to a decline in model performance; (3) Excessive calculation consumption. To address these issues, this paper proposes an Adaptive Hypergraph Contrastive Learning Network (AHCL-TC) for text classification, which combines graph contrastive learning and hypergraph neural networks to better capture the internal relationship structure of domain-specific terminology and achieve superior performance with imbalanced sample distribution. AHCL-TC designs a neural network structure based on hypergraphs, using the high-order relationships of hypergraphs to model complex structures in text data. This structure allows the model to better understand multiple relationships between terms, thereby improving classification performance. Graph contrastive learning is used as the training framework of the model. By learning the intrinsic characteristics and structure of graph data, and using data enhancement algorithms to expand the data set, the model's robustness to data imbalance is improved. Additionally, the paper presents a hypergraph adaptive augmentation algorithm designed for hypergraph structures to address the data imbalance problem. The proposed model is evaluated on multiple benchmark datasets, and the experimental results demonstrate its effectiveness for text classification tasks, outperforming baseline models even with reduced training data percentage. Furthermore, a comprehensive comparison of computational efficiency was conducted. The outcomes reveal that the computational consumption of our model is notably lower than that of other models. • The model can capture comprehensive semantic information and expand sample data. • This study introduces Hypergraph Attention Networks, leveraging hypergraph structures. • Comparative analysis shows both superior performance and computational efficiency. [ABSTRACT FROM AUTHOR]

Fine-grained image recognition is challenging because discriminative clues are usually fragmented, whether from a single image or multiple images. Despite their significant improvements, the majority of existing methods still focus on the most discriminative parts from a single image, ignoring informative details in other regions and lacking consideration of clues from other associated images. In this paper, we analyze the difficulties of fine-grained image recognition from a new perspective and propose a transformer architecture with the peak suppression module and knowledge guidance module, which respects the diversification of discriminative features in a single image and the aggregation of discriminative clues among multiple images. Specifically, the peak suppression module first utilizes a linear projection to convert the input image into sequential tokens. It then blocks the token based on the attention response generated by the transformer encoder. This module penalizes the attention to the most discriminative parts in the feature learning process, therefore, enhancing the information exploitation of the neglected regions. The knowledge guidance module compares the image-based representation generated from the peak suppression module with the learnable knowledge embedding set to obtain the knowledge response coefficients. Afterwards, it formalizes the knowledge learning as a classification problem using response coefficients as the classification scores. Knowledge embeddings and image-based representations are updated during training simultaneously so that the knowledge embedding includes a large number of discriminative clues for different images of the same category. Finally, we incorporate the acquired knowledge embeddings into the image-based representations as comprehensive representations, leading to significantly higher recognition performance. Extensive evaluations on the six popular datasets demonstrate the advantage of the proposed method in performance. The source code and models will be available online after the acceptance of the paper. [ABSTRACT FROM AUTHOR]

Point cloud analysis is a critical technology in the field of 3D vision, such as autonomous driving and robot navigation. Utilizing the inherent geometric properties embedded in 3D point cloud data remains a great challenge. In this paper, a cascaded geometric feature modulation network is proposed to explore the shared geometric patterns of 3D point clouds from local to global. The contribution of this paper is threefold. First, we design a local geometric feature modulation (GFM) block that learns a pointwise transformation according to the surrounding semantic context of each point. Second, based on the dense connection of multiple GFM blocks, we design a novel global fusion mechanism to ensure the preservation of valuable structural information. Finally, we propose a novel spatial distribution consistency loss to remedy the situation of irrational sampling. Benefitting from the proposed loss function, our network enjoys better convergence performance at the same time. Extensive experimental results on different tasks of point cloud processing demonstrate the superiority and robustness of our proposed network. [ABSTRACT FROM AUTHOR]

Point cloud classification for substation is challenging due to the occluding layout as well as the complexity of methods. Most previous methods trade inference speed for precision by building complex extractors. To alleviate the paradox of performance and complexity trade-off, this paper proposes the lightweight long-range attention classification for substation point cloud, including shuffle channel attention(SCA) and dilated channel attention(DCA). First, SCA captures local cross-channel interaction via 1D convolution and global interaction via channel shuffling, shown promising performance. Second, to further reduce the amount of computation involved in shuffling, we propose a more elegant method DCA by resizing the 1D vector after pooling into 2D feature map. Note that proposed DCA are implemented by just 2D convolution, determining the coverage of cross-channel interaction, with a significantly higher inference speed. Furthermore, we develop a heuristic algorithm to adaptively determine parameters like kernel size, shuffle group and size of 2D feature map. Besides ModelNet40 and ScanObjectNN(PB_T50_RS), this paper also selects ten of main objects in substation for training and testing. Finally, experimental results suggest that proposed methods bring notable performance gain, nearly 1%. DCA reaches 93.801% and just increases 0.0003M parameters but 15.5% faster than SCA. We also make an attempt to simplify proposed methods via reducing the coding dimensions and coding blocks. And simplified DCA achieves 93.4% performance with almost quarter parameters and 112% faster, guaranteeing both efficiency and effectiveness. A simplified version is used to achieve the highest accuracy of 85.115% and 83.304% on ScanObjectNN, with the highest improvements of 1.735% and 2.391%. This paper also conducts robustness test where different proportions of points are missing. Results show that when 93.75% missing, the accuracy only decreases by 4.093%, and can still reach 89%, which is significantly ahead of other methods. • We propose a high-fidelity point cloud style transfer module(PST), which transforms raw data into a unified Gaussian model while retaining its unique style to suit different feature extractors based on style deviation vector R d , scaling coefficient α and β. • This paper proposes a long-range attention method, called shuffle channel attention(SCA), which captures the local and the global cross-channel interaction in an efficient way. • To further improve the inference speed, we revisit SCA and propose the latest version: dilated channel attention(DCA). This method achieve a 15% advantage in speed via changing the shape of feature map, guaranteeing both efficiency and effectiveness. [ABSTRACT FROM AUTHOR]

Modern and new integrated technologies have changed the traditional systems by using more advanced machine learning, artificial intelligence methods, new generation standards, and smart and intelligent devices. The new integrated networks like the Internet of Things (IoT) and 5G standards offer various benefits and services. However, these networks have suffered from multiple object detection, localization, and classification issues. Conventional Neural Networks (CNN) and their variants have been adopted for object detection, classification, and localization in IoT networks to create autonomous devices to make decisions and perform tasks without human intervention and helpful to learn in-depth features. Motivated by these facts, this paper investigates existing object detection and recognition techniques by using CNN models used in IoT networks. This paper presents a Conventional Neural Networks for 5G-Enabled Internet of Things Network (CNN-5GIoT) model for moving and static objects in IoT networks after a detailed comparison. The proposed model is evaluated with existing models to check the accuracy of real-time tracking. The proposed model is more efficient for real-time object detection and recognition than conventional methods. [ABSTRACT FROM AUTHOR]

The existing Pavlov associative memory circuit only realizes the simple conditioned reflex process, secondary conditional reflex can make the simple conditioning process more complicated and make the circuit more bionic, but there is a lack of relevant circuit implementation. In this paper, a Pavlov associative memory circuit with secondary conditional reflex is proposed by utilizing the memristor. The proposed circuit can respond to a conditional stimulus after initial learning and have two kinds of forgetting process. Besides, this circuit can indirectly establish conditioned reflexes through conditioned stimuli, instead of directly establishing conditioned reflex with unconditioned stimulus. The realization of secondary conditional reflex is confirmed in the PSPICE simulation results. Meanwhile, an extended classification circuit based on secondary conditioned reflex is proposed. Based on the features of objects as input, the output of the circuit is used to achieve the function of classification. The accuracy of application circuit proposed in this paper can be verified by the simulation results in PSPICE. [ABSTRACT FROM AUTHOR]

• An analysis of lexicographic optimization for feature selection and data classification. • Cooperative co-evolution for feature selection while the classifier hyperparameters are also being optimized. • Optimization of multiple objectives using a simple evolutionary algorithm. • A new lexicographic ranking methodology to compare the results of different feature selection methods. • Quite good results for real high-dimensional datasets. This paper starts with two hypotheses. The first one is that the simultaneous optimization of the hyperparameters regulating the classifier within a wrapper method, while the best subset of features is being determined, should improve the results with respect to those obtained with a pre-parameterized classifier. The second one is that solving these two problems can be formulated as a lexicographic optimization problem, allowing the use of a simple single-objective evolutionary algorithm to solve this multi-objective problem. The fitness function is of key importance for such wrapper methods. It is responsible for guiding the search towards potentially good solutions and it also consumes most of the runtime. Having these issues in mind, this paper also proposes a new lexicographic fitness function, designed to minimize the runtime of the algorithm and also to avoid over-fitting. Furthermore, the execution time and the quality of the results obtained by the wrapper procedure also depend on some algorithmic hyperparameters: the similarity thresholds used when comparing two different solutions lexicographically and the percentage of data samples used for validation during the training process. Thus, an experimental analysis has been carried out to find adequate values for these hyperparameters. Finally, the lexicographic cooperative co-evolutionary wrapper approach, using the new fitness function proposed in this paper, has been tested with several datasets belonging to the University of California, Irvine (UCI) repository and also with some real high-dimensional datasets, obtaining quite good results, compared to other state-of-the-art wrapper methods. The comparison has also been made lexicographically, with a new methodology proposed in this paper. [ABSTRACT FROM AUTHOR]

With the rapid development of deep learning technology and improvement in computing capability, deep learning has been widely used in the field of hyperspectral image (HSI) classification. In general, deep learning models often contain many trainable parameters and require a massive number of labeled samples to achieve optimal performance. However, in regard to HSI classification, a large number of labeled samples is generally difficult to acquire due to the difficulty and time-consuming nature of manual labeling. Therefore, many research works focus on building a deep learning model for HSI classification with few labeled samples. In this article, we concentrate on this topic and provide a systematic review of the relevant literature. Specifically, the contributions of this paper are twofold. First, the research progress of related methods is categorized according to the learning paradigm, including transfer learning, active learning and few-shot learning. Second, a number of experiments with various state-of-the-art approaches has been carried out, and the results are summarized to reveal the potential research directions. More importantly, it is notable that although there is a vast gap between deep learning models (that usually need sufficient labeled samples) and the HSI scenario with few labeled samples, the issues of small-sample sets can be well characterized by fusion of deep learning methods and related techniques, such as transfer learning and a lightweight model. For reproducibility, the source codes of the methods assessed in the paper can be found at https://github.com/ShuGuoJ/HSI-Classification.git. [ABSTRACT FROM AUTHOR]

Mammography is an important screening criterion for breast cancer, one of the major diseases causing numerous deaths among female patients. Meanwhile, manual diagnosis of mammography is a time-consuming and labor-consuming job. Mammogram classification based on deep learning plays a vital role in computer-aided diagnosis (CAD) systems to mitigate the pressure on physicians. This paper proposes a learning-based multi-view mammogram classification model that captures long-distance dependence and extracts features of multiple receptive fields. Our model considers global and local features of mammography images using Transformer for global features and the proposed multiplex convolutions module for local features. We evaluate our proposed method on a dataset of mammography images obtained from a hospital in China. The proposed method achieves 90.57% accuracy and 94.86% AUC in benign or malignant classification tasks and outperforms other advanced methods for mammogram classification. It is worth noting that the proposed method only requires image-level labels and acts on the whole raw mammogram, which has clinical significance. [ABSTRACT FROM AUTHOR]

Multi-label classification associates each instance with a set of labels which reflects the nature of a wide range of real-world applications. However, existing approaches assume that all labels have the same misclassification cost, whereas in real-world problems different types of misclassification errors have different costs, which are generally unknown in the training context or might change from one context to another. Thus, there is a demand for cost-sensitive classification methods that minimise the average misclassification cost rather than error rates or counts. In this paper, we adopt a simple yet general method, called thresholding, which applies to most classification algorithms to adapt them to cost-sensitive multi-label classification. This paper investigates current threshold choice approaches for multi-label classification. It explores the choice of single and multiple thresholds and extends some of the current techniques to support multi-label problems. Moreover, it proposes cost curves and scatter diagrams for performance evaluation in the multi-label setting. Experimental evaluation on 13 multi-label datasets demonstrates that there is no significant loss by adjusting a global threshold rather than a per-label threshold considering different misclassification costs across labels. Although tuning multiple thresholds is the obvious solution, the global threshold can also be valid. [ABSTRACT FROM AUTHOR]

• We propose MetaGen: which can able to provide solutions for two tasks such as peer review prediction, meta-review generation. • Integrated framework of convolution layer, Bi-LSTM model, and attention mechanism to predict the final decision. • Provide a concise meta-review which maximizes information coverage, coherence, readability and also reduces redundancy. • Experiments on PeerRead dataset shows that MetaGen outperforms state-of-the-art methods. Peer reviews form an essential part of scientific communications. Research papers and proposals are reviewed by several peers before they are finally accepted or rejected for publication and funding, respectively. With the steady increase in the number of research domains, scholarly venues (journal and/or conference), researchers, and papers, managing the peer review process is becoming a daunting task. Application of recommender systems to assist peer reviewing is, therefore, being explored and becoming an emerging research area. In this paper, we present a deep learning network based Meta-Review Generation considering peer review prediction of the scholarly article (MRGen). MRGen is able to provide solutions for: (i) Peer review prediction (Task 1) and (ii) Meta-review generation (Task 2). First, the system takes the peer reviews as input and produces a draft meta-review. Then it employs an integrated framework of convolution layer, long short-term memory (LSTM) model, Bi-LSTM model, and attention mechanism to predict the final decision (accept/reject) of the scholarly article. Based on the final decision, the proposed model MRGen incorporates Pointer Generator Network-based abstractive summarization to generate the final meta-review. The focus of our approach is to give a concise meta-review that maximizes information coverage, coherence, readability and also reduces redundancy. Extensive experiments conducted on the PeerRead dataset demonstrate good consistency between the recommended decisions and original decisions. We also compare the performance of MRGen with some of the existing state-of-the-art multi-document summarization methods. The system also outperforms a few existing models based on accuracy, Rouge scores, readability, non-redundancy, and cohesion. [ABSTRACT FROM AUTHOR]

Classification problems where there exist multiple class variables that need to be jointly predicted are known as Multi-dimensional classification problems. If the labels of these class variables are organized as hierarchies, we can take advantage of specific strategies designed for the Hierarchical classification paradigm. In this paper we present the Multi-dimensional hierarchical classification (MDHC) paradigm, a result of the combination of Multi-dimensional and Hierarchical classification paradigms. We propose four MDHC learning strategies which are designed to exploit the particularities of this new paradigm, combining characteristics of Multi-dimensional and Hierarchical classification strategies. Along with these strategies, we present a framework for classifier comparison in which we use a set of performance measures specifically designed for MDHC, and a procedure to create MDHC synthetic scenarios. Using this framework and the performance measures presented, we study how characteristics of the MDHC problems influence the performance of the different MDHC strategies proposed, and compare them to other non-MDHC strategies. [ABSTRACT FROM AUTHOR]

Metric learning has attracted significant attention due to its high effectiveness and efficiency for pattern recognition task. Traditional supervised metric learning algorithms attempt to seek a global distance metric with labeled samples. When data are represented with multimodal and only limited supervision information is available, these approaches are insufficient to obtain satisfactory results. In this paper, we develop a robust semi-supervised multi-metric learning method (RSMM) to improve classification performance. The proposed RSMM learns multiple local metrics and a background metric instead of a single global metric. Specifically, we divide the metric space into influential regions and background region, and then regulate the effectiveness of each local metric to be within the related regions. Simultaneously, a geometrically interpretable, symmetric distance is defined with local metrics and background metric. Based on the resultant learning bounds, we obtain the regularization term to improve the classifier's generalization ability. Moreover, the manifold regularization term is introduced to preserve the supervision information as well as geometry structure. The substantial unlabeled samples may cause potential threats and large uncertainties, so the logarithmic loss function is utilized to enhance the robustness. An efficient gradient descent algorithm is exploited to solve the non-convex challenging problem. To further understand the proposed algorithm, we theoretically derive its robustness and generalization error bounds. Finally, numerical experiments on UCI datasets and image datasets demonstrate the feasibility and validity of the RSMM. [ABSTRACT FROM AUTHOR]

[Display omitted] • We propose a black-box adversarial text generation scheme called TextGuise, which can be used for almost all text classification tasks. The experiments show that TextGuise is superior to the state-of-the-art text adversarial example generation methods in terms of its attack success rate and attack efficiency when the perturbation ratio does not exceed 0.2.The adversarial examples generated using TextGuise retain the semantic information and strong readability of the text. None of the current defense methods can effectively defend against TextGuise. Adversarial examples greatly compromise the security of deep learning models. The key to improving the robustness of a natural language processing (NLP) model is to study attacks and defenses involving adversarial text. However, the current adversarial attack methods still face problems, such as the low success rates of attacks on some datasets, and the existing defense methods can already successfully defend against some attack methods. As a result, such attacks are unable to dig deeper into the flaws of NLP models to inform further defense improvements. Hence, it is necessary to design an adversarial attack method with a wider attack range and stronger performance. Aiming at the advantages and disadvantages of existing methods, this paper proposes a new adaptive black-box text adversarial example generation scheme, TextGuise. First, we design a keyword selection method in which word scores are calculated by combining context semantics to select the appropriate keywords to modify. Second, to maintain semantics, new keyword substitution rules are designed in combination with the characteristics of text and popular text expressions. Finally, the best modification strategy is adaptively selected through a querying model to reduce the magnitudes of disturbances. TextGuise can automatically select replacement keywords and replacement strategies that efficiently generate adversarial examples with good readability for various text classification tasks. Attack experiments conducted with TextGuise on 5 datasets yield high attack success rates that can surpass 80% when the perturbation ratio does not exceed 0.2. In addition, we present and discuss experiments focusing on defense, text similarity, query times, time consumption, etc., to test the attack performance of TextGuise. The results show that our attack method can achieve a good balance among various metrics. [ABSTRACT FROM AUTHOR]

Most Siamese trackers decompose the tracking task into two branches: classification and bounding box regression. They choose the bounding box with the highest classification score or the combination of the classification and predicted localization scores as the target. However, the misalignment between the classification and localization accuracy will degrade tracking performance. In this paper, we propose an IoU-aware Siamese tracker named IASNet, which predicts the IoU classification score of each regressed box to represent its localization confidence. We introduce a novel residual alignment module to capture the scale of each predicted bounding box and its border context, which further improves the reliability of the classification prediction. In addition, we build dynamic links between the classification and regression branches to enhance their interaction. The proposed dynamic loss enables the training to focus on high-quality positive samples, leading to more precise localization. Extensive experimental results on VOT2016, VOT2019, OTB100, GOT10k, UAV123, and LaSOT show that the proposed IASNet achieves state-of-the-art tracking performance and runs at approximately 65 fps on a GTX 3090 GPU, which far exceeds the real-time requirement. [ABSTRACT FROM AUTHOR]

Analysis of point clouds through deep convolutional neural networks is an active area of research due to their massive real-world applications including autonomous driving, indoor navigation, robotics, virtual/augmented reality, unmanned aerial vehicles, and drone technology. However, capturing the fine-grained geometric and semantic properties for the underlying recognition task with raw unstructured point cloud is highly challenging due to the lack of explicit neighborhood relationship and sparsity among the points. In this paper, we have introduced a deep, hierarchical 3D point based architecture for object classification and part segmentation that is able to learn robust geometric features which remain invariant to both the geometry and orientation of the local patches. The proposed architecture consists of multiple layers of sampling, concentric annular convolution, pooling, and residual feature propagation blocks. In the skip connections of our deep residual design, we propose to use a combination of linear projection shortcut and nonlinear ReLU group normalization shortcut with batch normalization, to improve both the optimization landscape and the representational power. Our network achieves on par or even better than state-of-the-art results on synthetic and real-world object classification (i.e., ModelNet40 and ScanObjectNN) and part segmentation (i.e., ShapeNet-part) benchmark datasets. The implementation and results have been made available at ...https://github.com/Rabbia-Hassan/Deep_Annular_Residual_Feature_Learning_for_3dPointCloudsGitHub-link [ABSTRACT FROM AUTHOR]

With the introduction of spatial-spectral fusion and deep learning, the classification performance of hyperspectral imagery (HSI) has been promoted greatly. For some widely used datasets, the classification accuracy almost reaches 100%. However, for hyperspectral image classification, random sampling is still the most common strategy to collect the training and test samples. Because the training and test samples are randomly selected from the same image, so they have a high correlation and the classification results are overoptimistic. Besides, random sampling is not a good choice for practical applications because we cannot always collect training and test samples from the same region. Disjoint sampling selects training and testing samples from different local regions, which will provide a more objective performance evaluation for HSI classification models. In this paper, we first show the huge classification performance difference caused by different sampling strategies with a simple experiment, then we analyze the underlying reasons from the spectral information, spatial-spectral combination, sample overlapping and spatial distance, finally, a semi-supervised feature learning method is proposed for disjoint HSI classification, in which the spatial and spectral information are exploited effectively and reasonably. The experimental results based on three HSI datasets demonstrate the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

The coordination of the body with the central nervous system has been studied using various biomechanical, neurophysiological, and neuroimaging studies. Different postural strategies provide evidence of cortical involvement to maintain postural stability, which can be utilised to minimise the risk of falls in the elderly and various pathological individuals. In this paper, we investigated the effect of vibrotactile feedback in Electroencephalography (EEG) based classification of voluntary postural sway during weight-shifting exercises in healthy and transfemoral amputees. The EEG data recorded during forward, backward, right, and left shifting as well as normal standing, with and without vibrotactile feedback, is decomposed using discrete wavelet transform. The energy of the coefficients from levels 4 to 7 forms the feature space to be forwarded to the weighted kNN classifier and ensemble bagged trees. We have achieved significantly higher classification rates across all the conditions for healthy and amputee subjects. Predictor importance from ensemble bagged tree models provides the highest contributions from the low-frequency band of 0–3.9 Hz and channels located over the motor and somatosensory cortex. We have also observed the contributions associated with the spinocerebellum and cerebrocerebellum. [ABSTRACT FROM AUTHOR]

Internet of Things (IoT) and wearable systems are very resource limited in terms of power, memory, bandwidth and processor performance. Sensor time series compression can be regarded as a direct way to use memory and bandwidth resources efficiently. On the other hand, the time series classification has recently attracted great attention and has found numerous potential uses in areas such as finance, industry and healthcare. This paper investigates the effect of lossy compression techniques on the time series classification task using deep neural networks. Furthermore, this paper proposes an efficient compression approach for univariate and multivariate time series that combines the lifting implementation of the discrete wavelet transform with an error-bound compressor, namely Squeeze (SZ), to attain an optimal trade-off between data compression and data quality. [ABSTRACT FROM AUTHOR]

Two new hybrid neural architectures combining morphological neurons and perceptrons are introduced in this paper. The first architecture, called Morphological - Linear Neural Network (MLNN) consists of a hidden layer of morphological neurons and an output layer of classical perceptrons has the capability of extracting features. The second architecture, called Linear-Morphological Neural Network (LMNN) is composed of one or several perceptron layers as a feature extractor, it is then followed by an output layer of morphological neurons for non-linear classification. Both architectures are trained by stochastic gradient descent. One of the main contributions of this paper is to show that the morphological layer offers a greater capacity to extract features than the perceptron layer. This claim is supported both theoretically and experimentally. We prove that the morphological layer possesses a greater capacity per computation unit to segment the 2D input space than the perceptron layer. In other words, adding more hyper-boxes produces more response regions than adding hyperplanes. From an empirical point of view, we test the two new models on 25 standard datasets at low dimensionality and one big data dataset. The result is that MLNN requires a lesser number of learning parameters than the other tested architectures while achieving better accuracies. [ABSTRACT FROM AUTHOR]

• This paper proposes a cofactorization method to perform jointly classification and representation learning. • The novel coupling approach produces a coherent and fully-interpretable hierarchical model. • The proposed model offers a rich interpretation of the data. • Performance is assessed on synthetic and real data in the specific context of hyperspectral image interpretation. Supervised classification and representation learning are two widely used classes of methods to analyze multivariate images. Although complementary, these methods have been scarcely considered jointly in a hierarchical modeling. In this paper, a method coupling these two approaches is designed using a matrix cofactorization formulation. Each task is modeled as a factorization matrix problem and a term relating both coding matrices is then introduced to drive an appropriate coupling. The link can be interpreted as a clustering operation over the low-dimensional representation vectors. The attribution vectors of the clustering are then used as features vectors for the classification task, i.e., the coding vectors of the corresponding factorization problem. A proximal gradient descent algorithm, ensuring convergence to a critical point of the objective function, is then derived to solve the resulting non-convex non-smooth optimization problem. An evaluation of the proposed method is finally conducted both on synthetic and real data in the specific context of hyperspectral image interpretation, unifying two standard analysis techniques, namely unmixing and classification. [ABSTRACT FROM AUTHOR]

• A scale-aware SPP module is proposed for selecting scale feature for each pixel. • An encoder-mask module is proposed to extract sharper boundary information. • An attention module is proposed to capture long-range dependencies. This paper focuses on semantic segmentation of scenes by capturing the appropriate scale, rich detail and contextual dependencies in a feature representation. Semantic segmentation is a pixel-level classification task and has made steady progress on the basis of fully convolutional networks (FCNs). However, we find there is still room for improvement in the following aspects. The first is that the pixel itself has not enough information for semantic prediction, it needs to look around to determine which category it belongs to. However, the fixed size of the receptive field defined by the network is not suitable for all pixels when an image contains objects with various scales. The second is that the extracted scale-aware features do not handle sharper object boundaries due to low-resolution. The final aspect is regarding the ability to model long-range dependencies. In order to solve the above challenges, in this paper we propose three modules: Scale-aware spatial pyramid pool module, Encoder mask module and Scale-attention module (SSPP-ES). Extensive experiments on the Cityscapes and ADE20K benchmarks demonstrate the effectiveness of our approach for semantic segmentation. [ABSTRACT FROM AUTHOR]

This paper presents an empirical exploration of the use of capsule networks for text classification. While it has been shown that capsule networks are effective for image classification, the research regarding their validity in the domain of text has been initiated recently. In this paper, we show that capsule networks indeed have the potential for text classification and that they have several advantages over convolutional neural networks. As well, we compare our proposed model to the initial studies regarding capsule network-based text classification. We further suggest a simple routing method that effectively reduces the computational complexity of dynamic routing. We utilized seven benchmark datasets to demonstrate that capsule networks, along with the proposed routing method provide comparable results. [ABSTRACT FROM AUTHOR]

• We propose a class of kernel interpolation-based methods, called kernel reconstruction learning, for solving machine learning problems. • We prove a reconstruction representer theorem, which indicates that conventional kernel methods can be viewed as special cases of kernel reconstruction learning. • We propose the kernel reconstruction vector machine, kernel reconstruction logistic regression, and kernel reconstruction density estimation methods, and show that they outperform popular kernel methods. This paper proposes a class of kernel interpolation-based methods, called kernel reconstruction learning, for solving machine learning problems. Kernel reconstruction learning uses kernel interpolators to reconstruct the unknown functions, which are needed to estimate in the problem, with estimated function values at selected knots. It can be applied to any learning problem that involves function estimation. We prove a reconstruction representer theorem, which indicates that conventional kernel methods, including kernel ridge regression, kernel support vector machine, and kernel logistic regression, can be viewed as special cases of kernel reconstruction learning. Furthermore, kernel reconstruction learning provides new algorithms for large datasets. The kernel reconstruction vector machine, kernel reconstruction logistic regression, and kernel reconstruction density estimation are discussed in detail. With appropriate implementations, they are shown to have higher prediction/estimation accuracy and/or less computational cost than popular kernel methods. [ABSTRACT FROM AUTHOR]

Data imbalance is an essential issue in multi-label image classification that may reduce the model's generalization ability. Unlike oversampling or undersampling in single-label image classification, simply removing or repeating some images in multi-label classification will influence labels in other categories due to complex label dependencies. Current methods usually tackle this issue from either data resampling or loss function design. This paper proposes a balanced masking strategy to solve the imbalance problem in feature space for graph-based methods. By selectively manipulating the label node embedding of the graph, the model can gain the priors of the majority of samples to maintain intra-class and inter-class balance simultaneously, thus performing better and more stable. Extensive experiments verify the adaptability and effectiveness of our method, which can be easily applied to any graph-based multi-label model to achieve highly competitive performance. [ABSTRACT FROM AUTHOR]

The problem's complexity assessment is an essential element of many topics in the supervised learning domain. It plays a significant role in meta-learning – becoming the basis for determining meta-attributes or multi-criteria optimization – allowing the evaluation of the training set resampling without needing to rebuild the recognition model. The tools currently available for the academic community, which would enable the calculation of problem complexity measures, are available only as libraries of the C++ and R languages. This paper describes the software module that allows for the estimation of 22 classification complexity measures and 12 regression complexity measures for the Python language – compatible with the scikit-learn programming interface – allowing for the implementation of research using them in the most popular programming environment of the machine learning community. [ABSTRACT FROM AUTHOR]

For imbalanced classification, data-level methods can achieve inter-class balance, but the samples generated do not contain new information and cannot avoid the problem of introducing noise. Algorithm-level methods may lead to overfitting of the model, and its classification effect is more dependent on the specific dataset and classification task, which means they lack universality. In addition, how to deeply mine the differences in the distribution of data overlap areas, and how to effectively mine the differences between categories when the absolute number of minority samples is small, are also important challenges in imbalanced classification. This paper proposes an imbalanced binary classification method using multi-label confidence comparisons based on contrastive learning. Different from the previous idea of directly learning its distribution characteristics from minority samples, combined with the idea of contrastive learning, the classification task is redefined as the multi-label matching task by mining the deep features that can represent the commonality and difference between the neighboring samples. Multiple differentiated contrastive sample groups are obtained through random sampling in its neighbor sample pool for each sample. This sample is combined with its contrastive sample groups to form multiple sample-neighbor pairs as training samples in the multi-label matching task. The original dataset is multiplied without introducing noise, laying a foundation for the effective mining of class differences when the absolute number of minority class samples is small. Based on the corresponding reconstruction error generated by Variational AutoEncoder (VAE), for sample-neighbor pairs, a multi-label matching loss between target samples and contrastive sample groups that integrates the idea of contrastive learning is designed. a robust classifier is obtained through simultaneous iterative learning of reconstruction error and multi-label matching loss, which can better mine the distribution differences of overlapping regions. In the testing phase, multiple different contrastive sample groups and the corresponding prediction results of the samples to be classified are obtained, which categories can be judged by integrating the predictions of each group for reverse reasoning. Experimental results on 38 public datasets show that the method outperforms typical imbalanced classification methods in both F1-measure and G-mean. [ABSTRACT FROM AUTHOR]

Although there have been a lot of researches on convolutional neural networks (CNNs), still what happens in this black box remains a mystery. In this paper, we establish the connection between CNNs and signal modulation. From a signal modulation point of view, the forward-propagation process of CNNs can be explained as a process of modulating the input signals to the vicinity of a special energy spectrum distribution, and the back-propagation process is searching for the appropriate distribution which is better for classification or other tasks. Several experiments have been carried out to verify the modulated explanation of CNNs. Furthermore, we verify that modulating the signal to the appropriate energy spectrum distribution in advance can effectively improve the classification and segmentation accuracy. [ABSTRACT FROM AUTHOR]

• GML (Group Meta-Learning) is proposed to effectively alleviate the problem that low-quality samples have unfavourable impact on the training effect under few-shot conditions, thereby improving the performance of the model. • Momentum update strategy is introduced to few-shot learning for the first time to effectively improve the generalization and stability of the model, and an adaptive momentum coefficient is designed to form AMS (Adaptive Momentum Smoothing) to further improve the training efficiency. • We propose MGML (Momentum Group Meta-Learning) by combining GML and AMS. MGML not only improves the accuracy of Meta-Learning Baseline, but also has good mobility. It can be inserted into previous state-of-the-art methods and get consistent performance improvements. At present, image classification covers more and more fields, and it is often difficult to obtain enough data for learning in some specific scenarios, such as medical fields, personalized customization of robots, etc. Few-shot image classification aims to quickly learn new classes of features from few images, and the meta-learning method has become the mainstream due to its good performance. However, the generalization ability of the meta-learning method is still poor and easy to be disturbed by low-quality images. In order to solve the above problems, this paper proposes Momentum Group Meta-Learning (MGML) to achieve a better effect of few-shot learning, which contains Group Meta-Learning module (GML) and Adaptive Momentum Smoothing module (AMS). GML obtains an ensemble model by training multiple episodes in parallel and then grouping them, which can reduce the interference of low-quality samples and improve the stability of meta-learning training. AMS includes the adaptive momentum update rule to further optimally integrate models between different groups, so that the model can memorize experience in more scenarios and enhance the generalization ability. We conduct experiments on miniImageNet and tieredImageNet datasets. The results show that MGML improves the accuracy, stability and cross-domain transfer ability of few-shot image classification, and can be applied to different few-shot learning models. [ABSTRACT FROM AUTHOR]

Time series classification is an increasingly attractive field with the appearance of new problems in an expanding digitalized world. Most of the proposals in the state-of-the-art have focused just on improving the results' performance, leaving interpretability on a secondary level. The available interpretable proposals do not provide competitive results, which is an issue to be addressed. This paper introduces a new fuzzy feature-based time series classification method, which joins the ability of time series features to capture essential information about the time series with Fuzzy logic. This proposal allows the fuzzy-based approach to incorporate global information about the behavior of time series in the membership calculation with the aim of improving the performance and interpretability of the results by using an interpretable classifier. The proposed method has been evaluated over the 112 state-of-the-art time series classification datasets from the UCR repository, and the results obtained show a better performance. Furthermore, the combination of time series features and fuzzy memberships has also increased the interpretability of final models. • PLAYTIME incorporates temporal information into fuzzy environments. • PLAYTIME combines fuzzy- and time-features to improve time series classification. • PLAYTIME improves interpretability and performance by using a tree-based classifier. • PLAYTIME is tested with 112 datasets of the UCR repository. [ABSTRACT FROM AUTHOR]

The co-occurrence of evolving concepts and imbalanced data deteriorates the learning performance of classifiers in data streams. Recent studies do not account for data difficulty factors associated with class imbalance, i.e. imbalance complexity, complicating the imbalance learning under a drifting data environment. This paper proposes EMRIL, a novel batch-based ensemble method, to deal with this challenge. As part of EMRIL, Imbalance Complexity Redressing Component (EMRIL ICRC), a data-level balancing module, resolves the imbalance complexity to increase minority class visibility for the base classifiers of the ensemble. Additionally, a novel ensemble pool management (EMRIL EPM) technique is designed using Reinforcement Learning (RL). EMRIL EPM regularly updates the ensemble pool and constructs an optimal base classifier subset for predictions through effective training and evaluation policies. Handling imbalance complexity, and RL-based ensemble pool management helps EMRIL to effectively perform the binary classification task in imbalanced and evolving data streams. A comprehensive experimental evaluation is conducted with 104 data streams which contain a variety of concept drifts and imbalance ratios categorized by various data difficulty factors. The results are compared with 15 state-of-the-art methods showing the superiority of the proposed method. [ABSTRACT FROM AUTHOR]

Given a text-image pair, Multimodal Named Entity Recognition (MNER) is the task of identifying and categorizing entities in the text. Most existing work performs named entity labeling directly using final token representations derived by fusing image and text representations. Although they achieve promising results, these work may fail to effectively exploit text and image modalities. This is because they neglect the difference in the role of the two modalities: text modality can detect the boundary of an entity, while image modality is introduced to disambiguate the category of the entity. Based on these findings, in this paper, we construct two auxiliary tasks based on the decomposition strategy and propose a multi-task framework for MNER. Specifically, we first decompose MNER into two auxiliary tasks: entity boundary detection task and entity category classification task. Here, the former treats only the text modality as input and outputs the boundary labels, since it can achieve satisfactory boundary results by itself. The latter uses two modalities to yield category labels where image modality is dedicated to disambiguating categories. These two auxiliary tasks allow the effective exploitation of text and image modalities and put them back into their respective roles. Then, we vectorize their results to improve entity recognition using label clues from auxiliary tasks. Finally, we fuse features from text and image modalities and label embeddings from auxiliary tasks to fulfill MNER. Experimental results on two widely used MNER datasets show that our framework can yield new SOTA performance. [ABSTRACT FROM AUTHOR]

Hierarchical classification is a significant challenge in computer vision due to the logical order and interconnectedness of multiple labels. This paper presents HD-CapsNet, a novel neural network architecture based on deep capsule networks, specifically designed for hierarchical multi-label classification(HMC). By incorporating a tree-like hierarchical structure, HD-CapsNet is designed to leverage the inherent ontological order within the hierarchical label tree, thereby ensuring classification consistency across different levels. Additionally, we introduce a specialized loss function that promotes accurate hierarchical relationships while penalizing inconsistencies. This not only enhances classification performance but also strengthens the network's robustness. We rigorously evaluate HD-CapsNet's efficacy by benchmarking it against existing HMC methods across six diverse datasets: Fashion-MNIST, Marine-Tree, CIFAR-10, CIFAR-100, Caltech-UCSD Birds-200-2011, and Stanford Cars. Our results conclusively demonstrate that HD-CapsNet excels in learning hierarchical relationships and significantly outperforms the competition in various image classification tasks. Our implementation is available at https://github.com/tasrif-khondaker/HD-CapsNet. [ABSTRACT FROM AUTHOR]

Partial Label Learning (PLL) is a weakly supervised learning problem that induces a multi-class classifier from data with candidate labels, among which only one is the ground-truth label. The crucial challenge in PLL is to disambiguate the false-positive labels from the candidate labels. However, most existing PLL methods fail to simultaneously consider both the instance-level similarity and the class-level information during label disambiguation. In this paper, we propose a novel two-stage method based on weighted centroid clustering, which efficiently utilizes both the local similarity among instances and the per-class global information. In the first stage, we initialize the center of each class using label propagation based on the instance-level similarity, and obtain the disambiguated instances via weighted centroid clustering derived from the per-class global information. In the second stage, the disambiguated instances are used to train a multi-class classifier. Extensive experiments on both controlled UCI datasets and real-world datasets show the superiority of the proposed method in classification accuracy. [ABSTRACT FROM AUTHOR]

Few-shot open-set recognition (FSOR) aims to develop models capable of generalizing to new tasks for both unknown detection and known classification with limited labeled data. Previous FSOR methods lack a comprehensive definition of the open space in the few-shot scenario, making models susceptible to overgeneralization. In this paper, we propose a novel method called Pairwise Discriminant Aggregation (PDAgg), addressing FSOR within a two-level recognition framework. PDAgg unifies the diverse optimization goals of FSOR at the pair level and provides a reasonable aggregate-level representation for unknown samples, thereby greatly enhancing model generalization to open space in the few-shot context. Specifically, PDAgg treats support-query pairs as the basic recognition units, which are adapted to a pair-specific feature space by enhancing pairwise representative features and incorporating a global knowledge context. Binary discriminant analyses are performed on adapted pair embeddings to estimate pair-level discriminant scores, which are then jointly aggregated to achieve both unknown detection and few-shot classification. Extensive experiments demonstrate that our method delivers comparable and even better performance with less extra information compared to the existing FSOR methods. • Pairwise Discriminant Aggregation (PDAgg) addresses FSOR with a two-level framework. • PDAgg offers a reasonable representation of unknown space in the few-shot context. • Extensive experiments showcase the superior performance of PDAgg for FSOR. [ABSTRACT FROM AUTHOR]

Artificial intelligence, and especially deep neural networks, have evolved substantially in the recent years, infiltrating numerous domains of applications, often greatly impactful to society's well-being. As a result, the need to understand how these models operate in depth and to access explanations of their decisions has become more vital than ever. Tending to this demand, the following paper aims to provide a thorough overview of the methods that have so far been developed to explain deep neural networks. Key aspects of explainability are defined and a straightforward classification of existing approaches is introduced, along with numerous examples. The task of realizing these methods on hardware is also discussed to complete the understanding of their application. [ABSTRACT FROM AUTHOR]

Graph convolutional networks (GCNs) have been very effective in addressing the issue of various graph-structured related tasks, such as node classification and graph classification. However, recent research has shown that GCNs are vulnerable to a new type of threat called a backdoor attack, where the adversary can inject a hidden backdoor into GCNs so that the attacked model performs well on benign samples, but its prediction will be maliciously changed to the attacker-specified target class if the hidden backdoor is activated by the attacker-defined trigger. A semantic backdoor attack is a new type of backdoor attack on deep neural networks (DNNs), where a naturally occurring semantic feature of samples can serve as a backdoor trigger such that the infected DNNs models will misclassify testing samples containing the predefined semantic feature even without the requirement of modifying the testing samples. Since the backdoor trigger is a naturally occurring semantic feature of the samples, semantic backdoor attacks are more imperceptible and pose a new and serious threat. Existing research on semantic backdoor attacks focuses on the tasks of CNNs-based (Convolutional Neural Networks) image classification and LSTM-based (Long Short-Term Memory) text classification or word prediction. Little attention has been given to semantic backdoor attacks on GCNs models. In this paper, we investigate whether such semantic backdoor attacks are possible for GCNs and propose a s emantic b ackdoor a ttack against G CNs (SBAG) under the context of graph classification to reveal the existence of this security vulnerability in GCNs. SBAG uses a certain type of nodes in the samples as a backdoor trigger and injects a hidden backdoor into GCNs models by poisoning training data. The backdoor will be activated, and the GCNs models will give malicious classification results specified by the attacker even on unmodified samples as long as the samples contain enough trigger nodes. We evaluate SBAG on five graph datasets. The experimental results indicate that SBAG can achieve attack success rates of approximately 99.9% on unmodified testing samples that naturally contain the trigger and attack success rates over 82% on testing samples modified to inject the trigger, respectively, both under poisoning rates of less than 5%. 1 1 Our code is available at: github. [ABSTRACT FROM AUTHOR]

Graph classification task plays a crucial role in many practical applications, for which the model is required to be capable of learning an accurate representation with discriminative characteristics, but existing methods typically fail to exploit the most authentic substructures for better representation. Aiming at this limitation, a fuzzy overlapping community guided subgraph neural network is proposed in this paper to make the best use of graph latent topologies for fine-grained representation learning and discriminative feature extraction, namely FS-GRAL. Taking multi-functional elements and their preferences fully into consideration, FS-GRAL allows the nodes to appear in more than one subgraph with different contributions for message passing, so that the model is powerful enough in capturing those natural aggregations inherent in complex networks to achieve accurate representation learning. Meanwhile, on the basis of learned accurate representations, a two-level pooling strategy dominated by intra-subgraph node discarding is presented for feature extraction, which enables graph latent topology can be integrated into extracted discriminative features, and thus the coarsened graph is more representative for graph classification. Extensive experiments on real-world benchmarks demonstrate that our fuzzy overlapping community guided subgraph neural network is highly competitive in learning the accurate graph representation and its intra-subgraph dominated two-level pooling strategy promotes a more promising classification performance. [ABSTRACT FROM AUTHOR]

The sparsity and redundancy of high-dimensional data poses challenges during processing. Dimensionality reduction is a commonly used approach, which helps to reduce data redundancy and noise, lower the complexity of learning algorithms, and improve classification accuracy. Typically, directly determining the intrinsic dimensionality of a dataset is challenging. Many of the present dimensionality reduction techniques suffer from pre-set target dimensions without considering the specific dataset features. To address this problem, a novel dimensionality reduction method based on flow model, which is called adaptive flow encoder (AFE), is proposed in this paper. Initially, the flow model is employed to achieve feature transformation and project the data onto an equip-dimensional feature space. Subsequently, a gating layer is utilized for feature selection. By configuring the gating layer parameters to be learnable, the model gains the ability to adaptively select features. To assess the efficacy of the proposed approach, the model's ability to preserve data characteristics is first qualitatively evaluated through visual inspection by projecting the data onto a two-dimensional space. Subsequently, five standard datasets are employed and compared against six traditional dimensionality reduction techniques to ascertain the impact on downstream classification tasks under the adaptive learning of intrinsic feature dimensions by the model. Moreover, the impact of key parameters, including the initialization scheme of the gate layer, amplification factor, and dimension factor, on the test results is examined. [ABSTRACT FROM AUTHOR]

Ordinal regression (OR) methods are designed for a type of classification problems where data labels have natural orders. In practice, data may be corrupted by label noise, which affects the training process thus degrading the generalization performance of OR methods. In OR, data are usually assumed to have latent variables underlying the ordinal labels, and label noise exhibits a special characteristic that it usually causes large latent variable value deviations. However, there are few existing works on OR considering label noise, and the existing works do not utilize the above-mentioned characteristic. Besides, most of the existing OR methods are centralized, which are inapplicable in some realistic distributed applications. In this paper, we utilize the characteristic of label noise in OR to develop a distributed robust support vector ordinal regression method (dRSVOR) under label noise. Specifically, after analyzing the characteristic of label noise in OR, we take the form of SVOR with explicit constraints to achieve robustness to one type of mislabeled samples. Then, we adopt correntropy, an information-theoretic measure, to achieve robustness to the other type of mislabeled samples. Theoretically, we analyze the consensus and convergence of dRSVOR. Experimentally, we conduct experiments on both synthetic data and real OR datasets to illustrate the effectiveness of the proposed method. The results show that the centralized version of dRSVOR outperforms several state-of-the-art OR methods considering label noise in centralized circumstances with label noise, and dRSVOR could approach the performance of the centralized version despite additional constraints in distributed scenarios. [ABSTRACT FROM AUTHOR]

Deep learning has enabled realistic face manipulation for malicious purposes (e.g., deepfakes), which poses significant concerns over the integrity of the media in circulation. Most existing deep learning techniques for deepfake detection can achieve promising performance in the intra-dataset evaluation setting, but are unable to perform satisfactorily in the inter-dataset evaluation setting. Most previous methods use a backbone network to extract global features for making predictions and only employ binary supervision to train the network. Classification merely based on the learning of global features often leads to weak generalizability to deepfakes of unseen manipulation methods. In this paper, we design a two-branch Convolutional AutoEncoder (CAE), which considers the reconstruction and classification tasks simultaneously for deepfake detection. This Joint Reconstruction and Classification (JRC) method shares the information learned by one task with the other, each focusing on different aspects, and hence boosts the overall performance. JRC is end-to-end, and experiments demonstrate that it achieves state-of-the-art performance on three commonly-used datasets, particularly in the cross-dataset evaluation setting. [ABSTRACT FROM AUTHOR]

Among the Graph Neural Network (GNN) models that address the task of node classification, Graph Echo State Networks (GESN) have proved particularly effective in addressing the challenge of heterophily, i.e. the presence of a significant fraction of inter-class edges in the learning task graph. The effectiveness of GESN is paired with its efficiency, owing to the reservoir computing paradigm. While previous literature has analyzed the design of reservoirs for sequence ESN and GESN for graph-level tasks, the problem of providing effective designs of reservoirs for node-level GESN is so far largely unexplored. In this paper we analyze the impact of different reservoir designs on node classification accuracy and on the quality of node embeddings computed by GESN, focusing both on dense and sparse reservoir layouts. As measures of embedding richness, we adopt both graph topology-dependent metrics previously employed in the analysis of embedding smoothing, and topology-independent metrics from the areas of information theory and numerical analysis. In particular, we propose the application of entropy measures for quantifying information in node embeddings. [Display omitted] • First systematic analysis of reservoir computing designs for node classification. • Proposing a set of topology-aware and -agnostic richness measures for node embeddings. • Random ensembles and dense orthogonal reservoirs perform generally better. • Extremely sparse reservoirs preserve accuracy and node embedding quality. [ABSTRACT FROM AUTHOR]

Pre-trained language models based on contrastive learning have shown to be effective in text classification. Although its great success, contrastive learning still has shortcomings. First, most of the existing contrastive learning methods neglect to consider linguistic knowledge, which can improve the performance of contrastive learning. Second, hard negatives have been proven to improve the quality of visual representations. However, generating hard negatives still needs to be explored in the field of text classification. In this paper, we propose a novel fine-grained (word-level) and coarse-grained (sentence-level) contrastive learning that can obtain high-quality sentence representations to improve the performance of text classification. Specifically, we construct word-level positive and negative sample pairs by WordNet and perform fine-grained contrastive learning to inject linguistic knowledge. Then we utilize the mixup to generate hard negatives while eliminating false negatives and use coarse-grained contrastive learning to train the model. Experiments on multiple public datasets demonstrate that our method outperforms state-of-the-art methods. [ABSTRACT FROM AUTHOR]

Feature selection has been recognized for long as an important preprocessing technique to reduce dimensionality and improve the performance of regression and classification tasks. The class of sequential forward feature selection methods based on Mutual Information (MI) is widely used in practice, mainly due to its computational efficiency and independence from the specific classifier. A recent work introduced a theoretical framework for this class of methods which explains the existing proposals as approximations to an optimal target objective function. Such framework made clear the advantages and drawbacks of each proposal. Methods accounting for the redundancy of candidate features using a maximization function and considering the so-called complementary effect are among the best ones. However, they still penalize the complementarity, which is an important drawback. This paper proposes the Decomposed Mutual Information Maximization (DMIM) method, which keeps the good theoretical properties of the best methods proposed so far but overcomes the complementarity penalization by applying the maximization separately to the inter-feature and class-relevant redundancies. DMIM was extensively evaluated and compared with other methods, both theoretically and using two synthetic scenarios and 20 publicly available real datasets applied to specific classifiers. Our results show that DMIM achieves a better classification performance than the remaining forward feature selection methods based on MI. [ABSTRACT FROM AUTHOR]

Local convolutional operation and global attention-based transformer operation can extract features of the point cloud from two different scales respectively, but few methods can combine them effectively. In this paper, we propose a novel point cloud feature extractor that has the advantages of both convolutional and transformer operation, which is named as CTpoint. Specifically, CTpoint is composed of two branches, where "C" represents the convolutional branch and "T" represents the transformer branch. The convolutional branch is responsible for extracting the local feature from grouped neighbor points, and the transformer branch performs the attention operation on the entire point cloud to capture the global feature. To make the two branches communicate with each other, the Feature Transmission Element (FTE) is proposed for the alignment of the spatial-feature dimension and the semantic space between the local and global features, so that the extracted local and global features can be effectively fused and the two branches can coordinately learn expressive features. We utilize CTpoint to construct point cloud classification and segmentation networks and evaluate their performance in several public datasets. In addition, we visualize the feature learned by CTpoint. The experimental results show that the expressive features learned by CTpoint make the networks achieve the state of the art performance on the point cloud classification and segmentation tasks. [ABSTRACT FROM AUTHOR]

Most text classification methods achieve great success based on the large-scale annotated data and the pre-trained language models. However, the labeled data is insufficient in practice, and the pre-trained language models are difficult to deploy due to their high computing resources and slow inference speed. In this paper, we propose cross-domain knowledge distillation, where the teacher and student tasks belong to different domains. It not only acquires knowledge from multiple teachers but also accelerates inference and reduces model size. Specifically, we train the pre-trained language models on factual knowledge obtained by aligned Wikipedia text to Wikidata triplets and fine-tune it as the teacher model. Then we use the heterogeneous multi-teacher knowledge distillation to transfer knowledge from the multiple teacher models to the student model. Multi-teacher knowledge vote can distill knowledge related to the target domain. Moreover, we also introduce the teacher assistant to help distill large pre-trained language models. Finally, we reduce the difference between the source domain and target domain by multi-source domain adaptation to solve the domain shift problem. Experiments on the multiple public datasets demonstrate that our method can achieve competitive performance while having fewer parameters and less inference time. [ABSTRACT FROM AUTHOR]

As 3D point cloud analysis has received increasing attention, the insufficient scale of point cloud datasets and the weak generalization ability of networks become prominent. In this paper, we propose a simple and effective augmentation method for the point cloud data to alleviate those problems. It finds the one-to-one correspondence between two point clouds and generates new training data by replacing the points in one sample with their corresponding pairs. Two replacement strategies are proposed to adapt to the accuracy or robustness requirement for different tasks, one of which is to randomly select all replacing points while the other one is to select k nearest neighbors of a single random point. Both strategies consistently and significantly improve the performance of various models on point cloud classification problems. By introducing the saliency maps to guide the selection of replacing points, the performance further improves. Since our method persists local semantic information of the points, we also at the first time extend MSDA for the point cloud segmentation problem. Moreover, PointCutMix is validated to enhance the model's robustness. When using as a defense method, our method outperforms the state-of-the-art defense algorithms. The code is available at: https://github.com/cuge1995/PointCutMix. [ABSTRACT FROM AUTHOR]

• Graph attention network (GAT) is a promising framework for message passing on graphs. • Increasing the attention range to more than one-hop neighbors can negatively affect the performance of GAT, reflecting the over-smoothing risk of graph neural networks in general. • The key idea is to contextualize each node with its "structural fingerprint" that can automatically adjust to the local graph topology and edge connections in the neighborhood of the node. • Our approach provides a useful platform for different subspace of node features and various spatial scale of graph structures to "cross-talk" with each other through multi-head attention, which is more flexible than existing attention mechanism using a fixed, minimal spatial attention scale. Graph attention network (GAT) is a promising framework for message passing on graphs, but how to exploit rich, high-order structural information in the attention mechanism is still an open challenge. Furthermore, increasing the attention range to more than one-hop neighbors can negatively affect the performance of GAT, reflecting the over-smoothing risk of graph neural networks in general. In this paper, we propose an "adaptive structural fingerprint" model to fully exploit complex graph topology in graph attention networks. The key idea is to contextualize each node with its "structural fingerprint" that can automatically adjust to the local graph topology and edge connections in the neighborhood of the node. By doing this, structural interactions between the nodes can be evaluated more accurately and better confined to relevant neighbors, thus contributing to an improved attention mechanism and clearer cluster boundary. Furthermore, our approach provides a useful platform for different subspace of node features and various spatial scale of graph structures to "cross-talk" with each other through multi-head attention, which is more flexible than existing attention mechanism using a fixed, minimal spatial attention scale. Encouraging results are observed on a number of benchmark data sets including citation and social networks. [ABSTRACT FROM AUTHOR]

Highlight • This paper studies semi-supervised TABSA via a deep generative model and achieves good performance. Abstract Target-oriented aspect-based sentiment analysis (TABSA) is a sentiment classification task that requires performing fine-grained semantical reasoning about a given aspect. The amount of labeled data is usually insufficient for supervised learning because the manual annotation w.r.t. the aspects is both time-consuming and laborious. In this paper, we propose a novel semi-supervised method to derive and utilize the underlying sentiment of unlabeled samples via a deep generative model. This method assumes that when given the aspect, the sentence is generated by two stochastic variables, i.e., the context variable and the sentiment variable. By explicitly disentangling the representation into the context and sentiment, the meaning of sentiment variable can be kept clean during the training phase. An additional advantage is that the proposed method uses a standalone classifier, and as such, our system is able to integrate with various supervised models. In terms of the implementation, since capturing the conditional input is non-trivial for a sequential model, special structures are put forward and investigated. We conducted experiments on SemEval 2014 task 4 and the results indicate that our method effectively handles five kinds of advanced classifiers. The proposed method outperforms two general semi-supervised methods and achieves state-of-the-art performance on this benchmark. [ABSTRACT FROM AUTHOR]