Neural Collapse Priors Driven Trust Semi-Supervised Multi-View Classification

Although deep neural networks have been proved effective in many applications, they are data hungry, and training deep models often requires laboriously labeled data. However, when labeled data contain erroneous labels, they often lead to model performance degradation. A common solution is to assign each sample with a dynamic weight during optimization, and the weight is adjusted in accordance with the loss. However, those weights are usually unreliable since they are measured by the losses of corrupted labels. Thus, this scheme might impede the discriminative ability of neural networks trained on noisy data. To address this issue, we propose a novel reweighting method, dubbed self-reweighting from class centroids (SRCC), by assigning sample weights based on the similarities between the samples and our online learned class centroids. Since we exploit statistical class centers in the image feature space to reweight data samples in learning, our method is robust to noise caused by corrupted labels. In addition, even after reweighting the noisy data, the decision boundaries might still suffer distortions. Thus, we leverage mixed inputs that are generated by linearly interpolating two random images and their labels to further regularize the boundaries. We employ the learned class centroids to evaluate the confidence of our generated mixed data via measuring feature similarities. During the network optimization, the class centroids are updated as more discriminative feature representations of original images are learned. In doing so, SRCC will generate more robust weighting coefficients for noisy and mixed data and facilitates our feature representation learning in return. Extensive experiments on both the synthetic and real image recognition tasks demonstrate that our method SRCC outperforms the state of the art on learning with noisy data.

Training large-scale Neural Networks requires substantial computational power and energy. Federated Learning enables distributed model training across geospatially distributed data centers, leveraging renewable energy sources to reduce the carbon footprint of AI training. Various client selection strategies have been developed to align the volatility of renewable energy with stable and fair model training in a federated system. However, due to the privacy-preserving nature of Federated Learning, the quality of data on client devices remains unknown, posing challenges for effective model training. In this paper, we introduce a modular approach on top to state-of-the-art client selection strategies for carbon-efficient Federated Learning. Our method enhances robustness by incorporating a noisy client data filtering, improving both model performance and sustainability in scenarios with unknown data quality. Additionally, we explore the impact of carbon budgets on model convergence, balancing efficiency and sustainability. Through extensive evaluations, we demonstrate that modern client selection strategies based on local client loss tend to select clients with noisy data, ultimately degrading model performance. To address this, we propose a gradient norm thresholding mechanism using probing rounds for more effective client selection and noise detection, contributing to the practical deployment of carbon-efficient Federated Learning.

The legibility of traffic signs has been considered from the beginning of design, and traffic signs are easy to identify for humans. For computer systems, however, identifying traffic signs still poses a challenging problem. Both image‐processing and machine‐learning algorithms are constantly improving, aimed at better solving this problem. However, with a dramatic increase in the number of traffic signs, labelling a large amount of training data means high cost. Therefore, how to use a small number of labelled traffic sign data reasonably to build an efficient and high‐quality traffic sign recognition (TSR) model in the Internet‐of‐things–based (IOT‐based) transport system has been an urgent research goal. Here, the authors propose a novel semi‐supervised learning approach combining global and local features for TSR in an IOT‐based transport system. In their approach, histograms of oriented gradient, colour histograms (CH), and edge features (EF) are used to build different feature spaces. Meanwhile, on the unlabelled samples, a fusion feature space is found to alleviate the differences between different feature spaces. Extensive evaluations on a collection of signs from the German Traffic Sign Recognition Benchmark (GTSRB) dataset shows that the proposed approach outperforms the others and provides a potential solution for practical applications.

Video-based face recognition has attracted a significant amount of research interest in both academia and industry due to its wide applications such as surveillance and security. Different from image-based face recognition, abundant information, extracted from a series of frames in a video, would contribute a lot to successful recognition. In other words, the key to improving video face recognition capability is aggregating and integrating profuse information within a video. Existing methods of feature aggregation across frames narrowly focus on the importance of a single frame, while ignoring the geometric relationship among frames in feature space. In this work, we present a geometry-based feature aggregation method rather than a better recognition model. It considers not only the importance of each frame but also the geometric relationship among frames in feature space, which yields more distinguishing video-level representation. Extensive evaluations on IJB-A and YTF datasets indicate that the proposed aggregation method considerably outperforms other feature aggregation methods.

Due to changes in global climate, coastal erosion is a common phenomenon, leading to flooding, habitat loss, property damage, and economic impacts. Large-scale automatic monitoring of coastlines is feasible by processing satellite images with Artificial Intelligence (AI) models; however, it is often constrained by scarce labels and model opacity. We present an explainable, fully unsupervised framework that first converts multi-year Sentinel-2 image series into pixel-wise land–water maps with attached model certainty and then discovers change patterns in the temporal structures of these land-water masks over a specified study period. Throughout the pipeline, we used the Normalized Difference Water Index (NDWI), computed from Sentinel-2 Green (B3) and NIR (B8), as our primary spectral indicator of surface water. We leverage two probabilistic models independently to select the best: (i) a Latent Dirichlet Allocation (LDA) model on a bag-of-visual-words to discover interpretable topics (e.g., foam, water, vegetation) that provide semantic segmentation with transparency, and (ii) a Gaussian Mixture Model (GMM) in pixel feature space to produce per-pixel probabilistic clustering and segmentation. Both models create multiple binary land-water segmentation maps from static images and then a single time-series evolution map from these multi-year land-water segmentation maps. The shoreline was automatically extracted as the boundary land pixels in binary maps with a threshold-based certainty score. We summarize changes with three methods and a score to define the amount of changes: the scene-level water fraction <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p_w</i>, shoreline-normal displacement (SND) along fixed transects, and the low-confidence proportion <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">U</i>τ that localizes ambiguity to physically dynamic interfaces. A Hamming score was computed to provide a quick and quantitative overview of the time-series evolution over a 7-year period. The pipeline delivers clear coastlines, interpretable segmentation maps, and uncertainty visualizations to help domain experts in auditing decisions. The result is a label-free, scalable, and explainable workflow in which every decision is accompanied by a confidence score, supporting reliable coastal monitoring and downstream scientific and application-oriented use.

Social networking sites such as Facebook or Twitter attract millions of users, who everyday post an enormous amount of content in the form of tweets, comments and posts. Since social network texts are usually short, learning tasks have to deal with a very high dimensional and sparse feature space, in which most features have low frequencies. As a result, extracting useful knowledge from such noisy data is a challenging task, that converts large-scale short-text learning tasks in social environments into one of the most relevant problems in machine learning and data mining. Feature selection is one of the most known and commonly used techniques for reducing the impact of the high dimensional feature space in text learning. A wide variety of feature selection techniques can be found in the literature applied to traditional, long-texts and document collections. However, short-texts coming from the social Web pose new challenges to this well-studied problem as texts’ shortness offers a limited context to extract enough statistical evidence about words relations (e.g. correlation), and instances usually arrive in continuous streams (e.g. Twitter timeline), so that the number of features and instances is unknown, among other problems. This paper surveys feature selection techniques for dealing with short texts in both offline and online settings. Then, open issues and research opportunities for performing online feature selection over social media data are discussed.

A new nonlinear dimensionality reduction method called data-dependent kernel global–local preserving projections (DDKGLPP) is proposed and used for process monitoring. To achieve performance improvements, DDKGLPP uses a data-dependent kernel rather than a conventional kernel. A unified kernel optimization framework is developed to optimize the data-dependent kernel by minimizing a data structure preserving index. The optimized kernel can unfold both global and local data structures in the feature space. The data-dependent kernel principal component (DDKPCA) and data-dependent kernel locality preserving projections (DDKLPP) also can be developed under the unified kernel optimization framework. However, unlike DDKPCA and DDKLPP, DDKGLPP is able to preserve both global and local structures of the data set when performing dimensionality reduction. Consequently, DDKGLPP is more powerful in capturing useful data characteristics. A DDKGLPP-based monitoring method is then proposed for nonlinear processes. Its per...

This paper presents a segregative genetic algorithm for optimizing the variable order in Reduced Ordered Binary Decision Diagrams. The main components are a basic genetic algorithm and two feature functions used to measure the similarity between chromosomes. Many copies of the basic genetic algorithm explore in parallel subpopulations induced in the search space by clustering in the feature space. A communication protocol preserves the similarity within each subpopulation during the evolution process. An associative tabu search memory is used to avoid reexploration of the search space. Extensive experimental evaluation proves the efficiency and stability of the segregative approach, which systematically produces better results than the basic genetic algorithm. The efficiency of the distributed implementation in terms of resource usage and many aspects regarding the communication protocol between different components are thoroughly described. The experiments used classical benchmarks known as very difficult and show that the segregative variant is better than the monopopulation algorithm and the approach using the island model.

Swarm intelligence based classifiers

Mass spectrometry-based metabolomics has undergone significant progresses in the past decade, with a variety of software packages being developed for data analysis. However, systematic comparison of different metabolomics software tools has rarely been conducted. In this study, several representative software packages were comparatively evaluated throughout the entire pipeline of metabolomics data analysis, including data processing, statistical analysis, feature selection, metabolite identification, pathway analysis, and classification model construction. LC-MS-based metabolomics was applied to preclinical Alzheimer’s disease (AD) using a small cohort of human cerebrospinal fluid (CSF) samples (N = 30). All three software packages, XCMS Online, SIEVE, and Compound Discoverer, provided consistent and reproducible data processing results. A hybrid method combining statistical test and support vector machine feature selection was employed to screen key metabolites, achieving a complementary selection of candidate biomarkers from three software packages. Machine learning classification using candidate biomarkers generated highly accurate and predictive models to classify patients into preclinical AD or control category. Overall, our study demonstrated a systematic evaluation of different MS-based metabolomics software packages for the entire data analysis pipeline which was applied to the candidate biomarker discovery of preclinical AD.

Ensemble data mining methods, also known as classifier combination, are often used to improve the performance of classification. Various classifier combination methods such as bagging, boosting, and random forest have been devised and have received considerable attention in the past. However, data dimensionality increases rapidly day by day. Such a trend poses various challenges as these methods are not suitable to directly apply to high-dimensional datasets. In this paper, we propose an ensemble method for classification of high-dimensional data, with each classifier constructed from a different set of features determined by partitioning of redundant features. In our method, the redundancy of features is considered to divide the original feature space. Then, each generated feature subset is trained by a support vector machine, and the results of each classifier are combined by majority voting. The efficiency and effectiveness of our method are demonstrated through comparisons with other ensemble techniques, and the results show that our method outperforms other methods.

Using a Feature Subset Selection method and Support Vector Machine to address curse of dimensionality and redundancy in Hyperion hyperspectral data classification

The quality of big data could great impact the value extracted from the data. Automated filtering of noisy data from big data is an ideal approach for improving the quality of big data. However, due to large volume and variety of big data, automated filtering of noisy data from big data is a grand challenging task. In this paper, we propose a support vector machine based approach for automated classification of big data so that the noisy data are classified as separated categories from the regular data. In order to improve the classification accuracy and training performance, we design an experiment for improving the classification model through finding the optimized learning feature set and an approach for iteratively improving the quality of the training data set. We conducted a thorough experimental study of automated classification of massive image data of biology cells to explain the approach of automated selection of big data and demonstrate its effectiveness. Finally, we compare the performance of the SVM based classification and a deep learning based classification of the same data set. The proposed approach and experience collected from the experimental study can help big data researchers and practitioners to design strategies for improving the quality of big data, designing high performance classifier, and building tools for automated selection of big data.

Fake news detection in low-resource languages: A novel hybrid summarization approach

Point clouds obtained with 3D scanners or by image‐based reconstruction techniques are often corrupted with significant amount of noise and outliers. Traditional methods for point cloud denoising largely rely on local surface fitting (e.g. jets or MLS surfaces), local or non‐local averaging or on statistical assumptions about the underlying noise model. In contrast, we develop a simple data‐driven method for removing outliers and reducing noise in unordered point clouds. We base our approach on a deep learning architecture adapted from PCPNet, which was recently proposed for estimating local 3D shape properties in point clouds. Our method first classifies and discards outlier samples, and then estimates correction vectors that project noisy points onto the original clean surfaces. The approach is efficient and robust to varying amounts of noise and outliers, while being able to handle large densely sampled point clouds. In our extensive evaluation, both on synthetic and real data, we show an increased robustness to strong noise levels compared to various state‐of‐the‐art methods, enabling accurate surface reconstruction from extremely noisy real data obtained by range scans. Finally, the simplicity and universality of our approach makes it very easy to integrate in any existing geometry processing pipeline. Both the code and pre‐trained networks can be found on the project page (https://github.com/mrakotosaon/pointcleannet).