Search

Due to advancements in data collection, storage, and processing techniques, machine learning has become a thriving and dominant paradigm. However, one of its main shortcomings is that the classical machine learning paradigm acts in isolation without utilizing the knowledge gained through learning from related tasks in the past. To circumvent this, the concept of Lifelong Machine Learning (LML) has been proposed, with the goal of mimicking how humans learn and acquire cognition. Human learning research has revealed that the brain connects previously learned information while learning new information from a single or small number of examples. Similarly, an LML system continually learns by storing and applying acquired information. Starting with an analysis of how the human brain learns, this paper shows that the LML framework shares a functional structure with the brain when it comes to solving new problems using previously learned information. It also provides a description of the LML framework, emphasizing its similarities to human brain learning. It also provides citation graph generation and scientometric analysis algorithms for the LML literatures, including information about the datasets and evaluation metrics that have been used in the empirical evaluation of LML systems. Finally, it presents outstanding issues and possible future research directions in the field of LML. [ABSTRACT FROM AUTHOR]

Owing to the predominant role of Machine Learning(ML) across domains, it is being introduced at multiple levels of education, including K-12. Researchers have leveraged games, augmented reality and other ways to make learning ML concepts interesting. However, most of the existing games to teach ML concepts either focus on use-cases and applications of ML instead of core concepts or directly introduce ML terminologies, which might be overwhelming to school students. Hence, in this paper, we propose ML-Quest, a game to incrementally present a conceptual overview of three ML concepts: Supervised Learning, Gradient Descent and K-Nearest Neighbor (KNN) Classification. The game has been evaluated through a controlled experiment, for its usefulness and player experience using the TAM model, with 41 higher-secondary school students. Results show that students in the experimental group perform better in the test than students in the control group, with 5% of students in the experimental group scoring full marks. However, none of the students in the control group could score full marks. The survey results indicate that around 77% of the participants who played the game either agree or strongly agree that ML-Quest has made their learning interactive and is helpful in introducing them to ML concepts. [ABSTRACT FROM AUTHOR]

Non-human-centered design tools, such as ChatGPT, have shown potential as effective aids in academic article design. This study conducts a comparative evaluation of ChatGPT-3.5 and ChatGPT-4, examining their capabilities and limitations in supporting the academic article design process. The study aims to demonstrate the utility of ChatGPT as a writing tool and investigate its applicability and efficacy in the context of academic paper design. The author interacted with both versions of ChatGPT, providing prompts and analyzing the generated responses. In addition, a different expert academic was consulted to assess the appropriateness of the ChatGPT responses. The findings suggest that ChatGPT, despite its limitations, could serve as a useful tool for academic writing, particularly in the design of academic articles. Despite the limitations of both GPT-3.5 and GPT-4, GPT-3.5 offers a broader perspective, whereas GPT-4 provides a more in-depth and detailed approach to the design of articles. ChatGPT exhibits capabilities in aiding the design process, generating ideas aligned with the overall purpose and focus of the paper, producing consistent and contextually relevant responses to various natural language inputs, partially assisting in literature reviews, supporting paper design in terms of both content and format, and providing reasonable editing and proofreading for articles. However, limitations were identified, including reduced critical thinking, potential for plagiarism, risk of misinformation, lack of originality and innovation, and limited access to literature. [ABSTRACT FROM AUTHOR]

With the development of smart classrooms, analyzing students' emotions for classroom learning is an effective means of accurately capturing their learning process. Although facial expression-based emotion analysis methods are effective in analyzing classroom learning emotions, current research focuses on facial expressions and does not consider the fact that expressions in different postures do not represent the same emotions. To provide a continuous and deeper understanding of students' learning emotions, this study proposes an algorithm to characterize learning emotions based on classroom time-series image data. First, face expression data for classroom scenarios are established to address the lack of expression databases in real teaching environments. Second, to improve the accuracy of facial expression recognition, a residual channel cross transformer masking net expression recognition model is proposed in this paper. Finally, to address the problem that the existing research dimension of learning emotion is too single, this paper uses the facial expression and head posture data obtained from deep learning models for fusion analysis and innovatively proposes a Dempster–Shafer evidence-theoretic fusion model to characterize the learning emotion within the lecture duration of knowledge points. The experiments show that both the proposed expression recognition model and the learning sentiment analysis algorithm have good performance, with the expression recognition model achieving an accuracy of 73.58% on the FER2013 dataset. The proposed learning emotion analysis method provides technical support for holistic analysis of student learning effects and evaluation of students' level of understanding of the knowledge points. [ABSTRACT FROM AUTHOR]

In the context of a full-scale war, participants in the educational process are implementing innovative strategies for adapting to new realities, based on the development and application of models of blended learning. The purpose of the academic paper is to study the features of the blended learning methodology in crisis situations with the identification of advantages and risks. The research was conducted using general scientific methods of cognition: analysis, synthesis, comparison, abstraction, specification, generalization, and formalization. In the course of the research, the existing models of the mixed type of educational process and the means of their practical implementation in educational institutions were analyzed. The problems of blended learning in the crisis period of military conflict are actualized, its features are presented. The risks accompanying the process of implementing the blended learning methodology are identified, and variations of measures to prevent them are proposed. It has been proven that the quality of the blended learning process depends on the teacher’s digital competencies in the online process. It has been established that the trend under study enriches the students’ learning experience and stimulates their motivation to study. The level of efficiency of blended learning is analyzed in comparison with distance and traditional learning. It has been determined that blended learning is much more effective than traditional and distance learning subject to careful planning of the educational process and a reasonable selection of tools. The essence of the methodology of blended learning as an innovative technology of the educational process and its effective anti-crisis format is proven. The practical significance of the research results lies in the possibility of their productive use in the process of optimizing the system of education in war conditions by improving existing and developing new methods and algorithms of blended learning. [ABSTRACT FROM AUTHOR]

In this paper, a novel Z-number based Fuzzy Neural Network (Z-FNN) based on the integration of Z-valued fuzzy logic and neural networks is proposed. Z-valued fuzzy rule base is presented and its inference process is described using interpolative approximate reasoning. Accordingly, the structure of the Z-FNN is proposed using a distance measure and interpolative approximate reasoning scheme. Based on presented architecture the learning algorithm of Z-FNN is designed. The updating of the unknown parameters of the network is carried out using Genetic Algorithms (GA). The proposed Z-FNN system is utilized for dynamic plant identification. The effectiveness of Z-FNN has been tested by comparing its performance with the performances of other fuzzy systems available in the literature. The proposed approach has been proven to be a suitable alternative for the identification of nonlinear systems characterized by uncertain and imprecise information. [ABSTRACT FROM AUTHOR]

• This paper proposes a novel multivariate dictionary learning framework. • A coherence enhancement term to improve the coherent representing of the coding coefficients under different resolutions. • A multivariate dictionary optimization method to solve dictionaries involving the calculation of fractional norm. • The proposed method achieves the state-of-the-art performance on several benchmark datasets. Off-the-shelf dictionary learning algorithms have achieved satisfactory results in small sample face recognition applications. However, the achieved results depend on the facial images obtained at a single resolution. In practice, the resolution of the images captured on the same target is different because of the different shooting equipment and different shooting distances. These images of the same category at different resolutions will pose a great challenge to these algorithms. In this paper, we propose a Joint Coupled Representation and Homogeneous Reconstruction (JCRHR) for multi-resolution small sample face recognition. In JCRHR, an analysis dictionary is introduced and combined with the synthetic dictionary for coupled representation learning, which better reveals the relationship between coding coefficients and samples. In addition, a coherence enhancement term is proposed to improve the coherent representation of the coding coefficients at different resolutions, which facilitates the reconstruction of the sample by its homogeneous atoms. Moreover, each sample at different resolutions is assigned a different coding coefficient in the multi-dictionary learning process, so that the learned dictionary is more in line with the actual situation. Furthermore, a regularization term based on the fractional norm is drawn into the dictionary coupled learning to remove the redundant information in the dictionary, which can reduce the negative impacts of the redundant information. Comprehensive results demonstrate that the proposed JCRHR method achieves better results than the state-of-the-art methods, on several small sample face databases. [ABSTRACT FROM AUTHOR]

Comparing the similarity of different enhancements of the same image is the key to achieving remarkable results in contrastive learning. The traditional contrastive learning method uses two different views of the image. In order to learn more information of the image to improve the classification accuracy, based on MoCo, this paper presented multi-view momentum contrast learning algorithms. In each iteration, this paper used a query encoder and multiple momentum encoders to extract feature for multiple data enhancements of the image, so that it could use more data enhancements and negative samples in this iteration. And it used the optimized noise contrast estimation(InfoNCE) to calculate the loss, so that the query encoder could obtain feature representations which was more beneficial to downstream tasks. The query encoder network updated with gradient backhaul, and each momentum encoder updated with an improved momentum update formula to improve the generalization ability of the model. Experimental results show that using multi-view momentum contrastive learning can effectively improve the classification accuracy of the model. [ABSTRACT FROM AUTHOR]

As an important component of the railway system, the surface damage that occurs on the rails due to daily operations can pose significant safety hazards. This paper proposes a simple yet effective rail surface defect detection model, FS-RSDD, for rail surface condition monitoring, which also aims to address the issue of insufficient defect samples faced by previous detection models. The model utilizes a pre-trained model to extract deep features of both normal rail samples and defect samples. Subsequently, an unsupervised learning method is employed to learn feature distributions and obtain a feature prototype memory bank. Using prototype learning techniques, FS-RSDD estimates the probability of a test sample belonging to a defect at each pixel based on the prototype memory bank. This approach overcomes the limitations of deep learning algorithms based on supervised learning techniques, which often suffer from insufficient training samples and low credibility in validation. FS-RSDD achieves high accuracy in defect detection and localization with only a small number of defect samples used for training. Surpassing benchmarked few-shot industrial defect detection algorithms, FS-RSDD achieves an ROC of 95.2% and 99.1% on RSDDS Type-I and Type-II rail defect data, respectively, and is on par with state-of-the-art unsupervised anomaly detection algorithms. [ABSTRACT FROM AUTHOR]

Background: Digital‐first assessments leverage the affordances of technology in all elements of the assessment process: from design and development to score reporting and evaluation to create test taker‐centric assessments. Objectives: The goal of this paper is to describe the engineering, machine learning, and psychometric processes and technologies of a test security framework (part of a larger ecosystem; Burstein et al., 2021) that can be used to create systems that protect the integrity of test scores. Methods: We use the Duolingo English Test to exemplify the processes and technologies that are presented. This includes methods for actively detecting and deterring malicious behaviour (e.g., a custom desktop app). It also includes methods for passively detecting and deterring malicious behaviour (e.g., a large item bank created through automatic generation methods). We describe the certification process that each test administration undergoes, which includes both automated and human review. Additionally, we describe our quality assurance dashboard which leverages psychometric data mining techniques to monitor test quality and inform decisions about item pool maintenance. Results and Conclusions: As assessment developers transition to online delivery and to a design approach that places the test taker at the centre, it becomes increasingly important to take advantage of the tools and methodological advances in different fields (e.g., engineering, machine learning, psychometrics). These tools and methods are essential to maintaining the security of assessments so that the score reliability is sustained and the interpretations and uses of test scores remain valid. Lay Description: What is known about this topic?: As more and more testing programmes transition to test taker‐centric administrations, effective measures to prevent cheating and protect content are critical to ensure the validity and integrity of scores.Two of the most common forms of cheating in online testing are (a) having someone other than the person who has registered take the test, and (b) stealing content and providing it to others to assist them in achieving a higher score. What does this paper add?: In designing a test taker‐centric digital‐first assessment, a test security framework must inform decisions from end‐to‐end (i.e., registration, onboarding, communications regarding test taker behaviours, test preparation and practice, test administration, and post‐administrative activities including scoring).Security is safeguarded through active and passive design methods; active methods include having test takers attest that they will follow the rules governing testing and informing them that they will be videoed during testing; passive methods include a computer adaptive design that limits item exposure and test overlap rates, development of a large item pool using automated item generation, and applying artificial intelligence to review test administration videos to flag unauthorized behaviours for human review. Implications: With more educational and assessment programmes transitioning to online digital models, the paper presents a comprehensive review of security issues and identifies an integrated approach for preventing cheating and other unauthorized behaviours. [ABSTRACT FROM AUTHOR]

Statistical machine learning usually achieves high-accuracy models by employing tens of thousands of examples. By contrast, both children and adult humans typically learn new concepts from either one or a small number of instances. The high data efficiency of human learning is not easily explained in terms of standard formal frameworks for machine learning, including Gold's learning-in-the-limit framework and Valiant's probably approximately correct (PAC) model. This paper explores ways in which this apparent disparity between human and machine learning can be reconciled by considering algorithms involving a preference for specificity combined with program minimality. It is shown how this can be efficiently enacted using hierarchical search based on identification of certificates and push-down automata to support hypothesizing compactly expressed maximal efficiency algorithms. Early results of a new system called DeepLog indicate that such approaches can support efficient top-down construction of relatively complex logic programs from a single example. This article is part of a discussion meeting issue 'Cognitive artificial intelligence'. [ABSTRACT FROM AUTHOR]

In a society centered on hyper-connectivity, information sharing is crucial, but it must be ensured that each piece of information is viewed only by legitimate users; for this purpose, the medium that connects information and users must be able to identify illegal users. In this paper, we propose a smartphone authentication system based on human gait, breaking away from the traditional authentication method of using the smartphone as the medium. After learning human gait features with a convolutional neural network deep learning model, it is mounted on a smartphone to determine whether the user is a legitimate user by walking for 1.8 s while carrying the smartphone. The accuracy, precision, recall, and F1-score were measured as evaluation indicators of the proposed model. These measures all achieved an average of at least 90%. The analysis results show that the proposed system has high reliability. Therefore, this study demonstrates the possibility of using human gait as a new user authentication method. In addition, compared to our previous studies, the gait data collection time for user authentication of the proposed model was reduced from 7 to 1.8 s. This reduction signifies an approximately four-fold performance enhancement through the implementation of filtering techniques and confirms that gait data collected over a short period of time can be used for user authentication. [ABSTRACT FROM AUTHOR]

Prediction of software bugs from process logs, temporal access logs, behavior analysis, etc. requires estimation of a wide variety of high-density feature sets. Extracted feature sets must be able to classify these logs into different bug categories with high accuracy, and low complexity. To perform these tasks, a wide variety of Machine Learning Models (MLMs) are proposed by researchers, and each of them varies in terms of their performancelevel nuances, functional advantages, contextual limitations, and application-specific future scopes. Upon analyzing these characteristics, it was observed that existing models are highly context-specific, and cannot be applied to multidomain bug analysis datasets. Moreover, existing models do not incorporate a dynamic feature selection method, which limits their accuracy performance under multiple bug classification applications. To overcome these issues, this paper proposes design of a novel Ensemble deep learning Classifier with Feature Selection for highefficiency Multidomain Bug Prediction under different use cases. The proposed model improves bug representation performance by combining multiple feature extraction methods, including GWO-based novel feature selection techniques. The ensemble classification model, which combines Deep Random Forest, k Nearest Neighbor, Logistic Regression, Multilayer Perceptron, Support Vector Machine, and 1D Convolutional Neural Network classifiers, achieves higher accuracy, precision, recall, and low delay compared to existing models. The model also shows faster classification speeds than existing models and can be deployed for various real-time applications. This accuracy performance was compared with various state-of-the-art models, and it was observed that the proposed model showcases 4.5% higher accuracy, 3.2% better precision, 3.9% higher recall, and 5.5% faster classification performance, which was possible due to integration of intelligent feature selection process with high efficiency classification models under multidomain scenarios. [ABSTRACT FROM AUTHOR]

Hierarchical temporal memory is an emerging machine learning technology that aims to model the structural and algorithmic properties of the neocortex. It is particularly suitable for learning and predicting sequential data. However, when dealing with long time series or complex sequences, the accuracy is relatively lower than desired. In this paper, a novel hierarchical temporal memory based on recurrent learning unit is proposed, where a feedback mechanism is involved into the model. The original cell is extended with a recurrent unit to capture long temporal dependencies of synaptic connections between neurons. The temporal pooler algorithm is then modified to adapt to the recurrent learning unit, and the supervised gradient information is combined with the Hebbian synaptogenesis learning rule in speeding up the training. The prototype of the proposed hierarchical temporal memory is implemented and extensive experiments are carried out on two public datasets under various settings. Experimental results show that the proposed model obtains an accuracy increase by up to 32% and a perplexity drop by up to 14% on sequence prediction and text generation tasks, respectively, which indicates the hierarchical temporal memory with recurrent feedback outperforms the original model on sequence learning. [ABSTRACT FROM AUTHOR]

Getting stuck is an inevitable part of learning programming. Long-term stuck decreases the learner's motivation and learning efficiency. The current approach to supporting learning in lectures involves teachers finding students who are getting stuck, reviewing their source code, and solving the problems. However, it is difficult for teachers to grasp every learner's stuck situation and to distinguish stuck or deep thinking only by their source code. Teachers should advise learners only when there is no progress and they are psychologically stuck. This paper proposes a method for detecting when learners get stuck during programming by using multi-modal data, considering both their source code and psychological state measured by a heart rate sensor. The evaluation results of the proposed method show that it can detect more stuck situations than the method that uses only a single indicator. Furthermore, we implemented a system that aggregates the stuck situation detected by the proposed method and presents them to a teacher. In evaluations during the actual programming lecture, participants rated the notification timing of application as suitable and commented that the application was useful. The questionnaire survey showed that the application can detect situations where learners cannot find solutions to exercise problems or express them in programming. [ABSTRACT FROM AUTHOR]

Educational Data Mining (EDM) has become a promising research field for improving the quality of students and the education system. Although EDM dates back to several years, there is still lack of works for measuring and enhancing the computer programming skills of tertiary students. As such, we, in this paper, propose an EDM system for evaluating and improving tertiary students' programming skills. The proposed EDM system comprises two key modules for (i) classification process and (ii) learning process,. The classification module predicts the current status of a student and the learning process module helps generate respective suggestions and feedback to enhance the student's quality. In particular, for the classification module, we prepare a real dataset related to this task and evaluate the dataset to investigate six key Machine Learning (ML) algorithms, Support Vector Machine (SVM), decision tree, artificial neural network, Random Forest (RF), k -nearest neighbor and naive Bayes classifier, using accuracy-related performance measure metrics and goodness of the fit. The experimental results manifest that RF and SVM can predict the students more accurately than the other models. In addition, critical factors analysis is accomplished to identify the critical features toward achieving high classification accuracy. At last, we design an improvement mechanism in the learning process module that helps the students enhance their programming skills. [ABSTRACT FROM AUTHOR]

The idea of EAE (Ecological Aesthetic Education) is put forward on the basis of the mature development of ecological aesthetics and AE theory. Starting with EAE, establishing people's aesthetic attitude and improving people's spiritual realm will help to reverse people's hostile attitude towards nature and rebuild the harmonious relationship between man and nature. This paper studies the application optimization of AE (aesthetic education) resources in universities based on ML (machine learning) from the perspective of resource development. The recommendation algorithm based on ML is the main idea of the classification of AE resources, and the classification model of AE resources is constructed. Through deep learning, we can learn the effective features of items from the content data in advance and then transform the learned features into CF (Collaborative Filtering) target learning task. Optimize the voting mechanism of the algorithm, and compare the RF_VM (random forest with optimized voting mechanism) algorithm with the traditional RF (random forest) algorithm. Experiments show that the algorithm proposed in this paper can effectively classify texts and has high feasibility. [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]

Class imbalance distribution widely exists in real-world engineering. However, the mainstream optimization algorithms that seek to minimize error will trap the deep learning model in sub-optimums when facing extreme class imbalance. It seriously harms the classification precision, especially in the minor classes. The essential reason is that the gradients of the classifier weights are imbalanced among the components from different classes. In this paper, we propose Attraction–Repulsion-Balanced Loss (ARB-Loss) to balance the different components of the gradients. We perform experiments on large-scale classification and segmentation datasets, and our ARB-Loss can achieve state-of-the-art performance via only one-stage training instead of 2-stage learning like nowadays SOTA works. [ABSTRACT FROM AUTHOR]

Cost-sensitive learning (CSL), which has gained widespread attention in class imbalance learning (CIL), can be implemented either by tuning penalty parameters or by designing new loss functions. In this paper, we propose a cost-sensitive learning method with a modified Stein loss function (CSMS) and a robust CSMS (RCSMS). Specifically, CSMS is flexible, as it realizes CSL from above two aspects simultaneously. In contrast, RCSMS merely achieves CSL by tuning penalty parameters, but the adopted loss function makes it insensitive to noise. To our best knowledge, it is the first time for Stein loss function derived from statistics to be applied in machine learning, which not only offers two alternative class imbalance solutions but also provides a novel idea for the design of loss functions in CIL. The mini-batch stochastic sub-gradient descent (MBGD) approach is employed to optimize CSMS and RCSMS. Meanwhile, the Rademacher complexity is used to analyze their generalization error bounds. Extensive experiments profoundly confirm the superiority of both models over benchmarks. [ABSTRACT FROM AUTHOR]

Federated learning is designed to collaboratively train a shared model based on a large number of mobile devices while preserving data privacy, which has been widely adopted to support different geo-spatial systems. However, two critical issues prevent federated learning to be effectively deployed on resource-constrained devices in large scale. First, federated learning causes high energy consumption which can badly hurt the battery lifetime of mobile devices. Second, leakage of sensitive personal information still occurs during the training process. Thus, a system that can effectively protect the sensitive information while improving the energy efficiency is urgently required for a mobile-based federated learning system. This paper proposes SmartDL, an energy-aware decremental learning framework that well balances the energy efficiency and data privacy in an efficient manner. SmartDL improves the energy efficiency from two levels: (1) global layer, which adopts an optimization approach to select a subset of participating devices with sufficient capacity and maximum reward. (2) local layer, which adopts a novel decremental learning algorithm to actively provides the decremental and incremental updates, and can adaptively tune the local DVFS at the same time. We prototyped SmartDL on physical testbed and evaluated its performance using several learning benchmarks with real-world traces. The evaluation results show that compared with the original federated learning, SmartDL can reduce energy consumption by 75.6–82.4% in different datasets. Moreover, SmartDL achieves a speedup of 2–4 orders of magnitude in model convergence while ensuring the accuracy of the model. [ABSTRACT FROM AUTHOR]

Large-scale geospatial data pave the way for geospatial machine learning algorithms, and a good representation is related to whether the machine learning model is effective. Hence, it is a critical task to learn effective feature representation for geospatial data. In this paper, we construct a spatial graph from the locations and propose a geospatial graph contrastive learning method to learn the location representations. Firstly, we propose a skeleton graph in order to preserve the primary structure of the geospatial graph to solve the positioning bias problem of remote sensing. Then, we define a novel mixed node centrality measure and propose four data augmentation methods based on the measure. Finally, we propose a heterogeneous graph attention network to aggregate information from both the structural neighborhood and semantic neighborhood separately. Extensive experiments on both geospatial datasets and non-geospatial datasets are conducted to illustrate that the proposed method outperforms state-of-the-art baselines. [ABSTRACT FROM AUTHOR]

In recent years, the rapid development of deep learning approaches has paved the way to explore the underlying factors that explain the data. In particular, several methods have been proposed to learn to identify and disentangle these underlying explanatory factors in order to improve the learning process and model generalization. However, extracting this representation with little or no supervision remains a key challenge in machine learning. In this paper, we provide a theoretical outlook on recent advances in the field of unsupervised representation learning with a focus on auto-encoding-based approaches and on the most well-known supervised disentanglement metrics. We cover the current state-of-the-art methods for learning disentangled representation in an unsupervised manner while pointing out the connection between each method and its added value on disentanglement. Further, we discuss how to quantify disentanglement and present an in-depth analysis of associated metrics. We conclude by carrying out a comparative evaluation of these metrics according to three criteria, (i) modularity, (ii) compactness and (iii) informativeness. Finally, we show that only the Mutual Information Gap score (MIG) meets all three criteria. [ABSTRACT FROM AUTHOR]

Accurate estimation of State of Health (SOH) is crucial to ensure optimal performance and safe operation of lithium-ion battery. This paper proposes a Stacking ensemble learning paradigm for SOH estimation. The Stacking ensemble learning increases adaptability to different features by using base learners with different structures, reducing the risk of overfitting. The model utilizes random vector functional link (RVFL) and active state tracking long-short-term memory network (AST-LSTM) as base learners, where AST-LSTM actively tracks long-term information of lithium-ion battery, and RVFL acts as the meta-learner for stacking. The random vector functional link network helps to avoid the problem of gradient vanishing that is commonly encountered in neural networks due to the gradient descent principle. To further improve estimation accuracy, Singer initialization method and dimension learning method are employed to enhance the Heap-based optimization (HBO) algorithm. In this study, the IHBO algorithm is used to optimize the hyperparameters of the model. Comparing with other methods, the hybrid model proposed in this paper demonstrates superior estimation performance under different operating conditions: at a temperature of 24 °C with a discharge current of 1 A, at a temperature of 4 °C with a discharge current of 1 A, and at a temperature of 4 °C with a discharge current of 2 A. The highest RMSE of the proposed method for the three working conditions are 0.006, 0.01, and 0.017, respectively. Therefore, the proposed Stacking ensemble learning is feasible for SOH estimation of lithium-ion battery and can better adapt to lithium-ion battery data under different operating conditions. • A Stacking ensmeble learning strategy is proposed for SOH estimation. • The improved HBO algorithm is introduced to optimize the Stacking model. • SOH for lithium batteries under three different operating conditions is estimated. • Six benchmark models are used to verify the performance of the proposed model. [ABSTRACT FROM AUTHOR]

Navigation is one of the most heavily studied problems in robotics and is conventionally approached as a geometric mapping and planning problem. However, real-world navigation presents a complex set of physical challenges that defies simple geometric abstractions. Machine learning offers a promising way to go beyond geometry and conventional planning, allowing for navigational systems that make decisions based on actual prior experience. Such systems can reason about traversability in ways that go beyond geometry, accounting for the physical outcomes of their actions and exploiting patterns in real-world environments. They can also improve as more data is collected, potentially providing a powerful network effect. In this article, we present a general toolkit for experiential learning of robotic navigation skills that unifies several recent approaches, describe the underlying design principles, summarize experimental results from several of our recent papers, and discuss open problems and directions for future work. This article is part of the theme issue 'New approaches to 3D vision'. [ABSTRACT FROM AUTHOR]

Sparse unmixing is a semisupervised unmixing method based on the linear mixture model, in which the spectral library is known a prior, and has received considerable attention recently. It has been confirmed that the spatial information in hyperspectral images plays a crucial role in improving the performance of sparse unmixing algorithms. However, the spatial information extracted or captured in most unmixing algorithms is inaccurate and robust enough, which leads to artificial block noise or outliers in the estimated abundance maps, especially as the noise level increases. To address these problems and more efficiently utilize the spatial information, this paper proposes a graph learning and denoising-based weighted sparse mixing (GLDWSU) algorithm, which includes three stages in the unmixing procedure: graph learning, denoising and unmixing. In the first stage, the graph Laplacian matrix is adaptively learned to capture the spatial structure of HSI with the relative total variation (RTV) regularization. In the next stage, HSI is denoised with the learned Laplacian matrix by using Laplacian smoothing. In the final stage, the denoised HSI is unmixed with the reweighted-norm regularization based on the alternating direction method of multipliers (ADMM) framework. The experimental results on both simulated and real data sets show that the proposed GLDWSU algorithm can more accurately capture and utilize the spatial structure information of HSI with a low computational cost and outperforms all the compared methods. [ABSTRACT FROM AUTHOR]

We present a novel online decision-making solution, where the optimal path of a given decision tree is dynamically found based on the contextual bandits analysis. At each round, the learner finds a path in the decision tree by making a sequence of decisions following the tree structure and receives an outcome when a terminal node is reached. At each decision node, the environment information is observed to hint on which child node to visit, resulting in a better outcome. The objective is to learn the context-specific optimal decision for each decision node to maximize the accumulated outcome. In this paper, we propose Dynamic Path Identifier (DPI), a learning algorithm where the contextual bandit is applied to every decision node, and the observed outcome is used as the reward of the previous decisions of the same round. The technical difficulty of DPI is the high exploration challenge caused by the width (i.e., the number of paths) of the tree as well as the large context space. We mathematically prove that DPI's regret per round approached zero as the number of the rounds approaches infinity. We also prove that the regret is not a function of the number of paths in the tree. Numerical evaluations are provided to complement the theoretical analysis. [ABSTRACT FROM AUTHOR]

Federated Learning is a novel framework that allows multiple devices or institutions to train a machine learning model collaboratively while preserving their data private. This decentralized approach is prone to suffer the consequences of data statistical heterogeneity, both across the different entities and over time, which may lead to a lack of convergence. To avoid such issues, different methods have been proposed in the past few years. However, data may be heterogeneous in lots of different ways, and current proposals do not always determine the kind of heterogeneity they are considering. In this work, we formally classify data statistical heterogeneity and review the most remarkable learning Federated Learning strategies that are able to face it. At the same time, we introduce approaches from other machine learning frameworks. In particular, Continual Learning strategies are worthy of special attention, since they are able to handle habitual kinds of data heterogeneity. Throughout this paper, we present many methods that could be easily adapted to the Federated Learning settings to improve its performance. Apart from theoretically discussing the negative impact of data heterogeneity, we examine it and show some empirical results using different types of non-IID data. • We review the advances in Federated Learning, particularly on heterogeneous data. • We consider time-evolving situations in Federated Learning. • We establish a classification of the literature according to their data heterogeneity. • We settle restrictions inherent to some of the problem scenarios. • We examine the open challenges and future directions for Federated Learning. [ABSTRACT FROM AUTHOR]

Sensor-based human activity recognition has been extensively studied. Systems learn from a set of training samples to classify actions into a pre-defined set of ground truth activities. However, human behaviours vary over time, and so a recognition system should ideally be able to continuously learn and adapt, while retaining the knowledge of previously learned activities, and without failing to highlight novel, and therefore potentially risky, behaviours. In this paper, we propose a method based on compression that can incrementally learn new behaviours, while retaining prior knowledge. Evaluation was conducted on three publicly available smart home datasets. [ABSTRACT FROM AUTHOR]

Human health diseases are caused due to genetic disorders, environmental conditions, presence of Alpha synuclein bodies, etc. Identification of this disease can be done via estimation of body tremors, slowed movements, fever, change in corpuscles, rigid muscles, impaired posture & balance, pain analysis, loss of automatic movements, speech changes, and writing changes. To estimate these effects, researchers have proposed a wide variety of machine learning techniques that can analyze speech signals, electroencephalogram (EEG) signals, electrocardiogram (ECG) signals, handwriting variations, gait identification & classification, etc. These models are highly variant in terms of their internal operating characteristics, which makes it ambiguous for clinical designers to identify optimal models for their deployments. Furthermore, these models also showcase highly variant performance levels as per their input feature sets. Thus, it is difficult to identify optimal models for a given set of input parameters. To overcome these issues, this paper initially reviews a wide variety of Human Health disease identification models in terms of their contextual nuances, operating advantages, functional limitations, quantitative characteristics, and deployment-specific future scopes. This will assist readers to identify functional models that can be applied for their operation-specific use cases. After referring this comparison, readers will be presented with a comparative analysis of these models in terms of their detection accuracy, classification delay, input datasets, computational complexity, and scalability levels. Based on this comparison, readers will be able to identify optimally performing models as per their qualitative characteristics. [ABSTRACT FROM AUTHOR]

As discarding superfluous instances in data sets shortens the learning process, it also increases learning performance because of eliminating noisy data. Instance selection methods are commonly utilized to undertake the abovementioned tasks. In this paper, we propose a new supervised instance selection algorithm called Border Instances Reduction using Classes Handily (BIRCH). BIRCH considers k-nearest neighbors of each instance and selects instances that have neighbors from the only same class, namely, but not having neighbors from the different classes. It has been compared with one traditional and four state-of-the-art instance selection algorithms by using fifteen data sets from various domains. The empirical results show BIRCH well delivers the trade-off between accuracy rate and reduction rate by tuning the number of neighbors. Furthermore, the proposed method guarantees to yield a high classification accuracy. The source code of the proposed algorithm can be found in https://github.com/fatihaydin1/BIRCH. [ABSTRACT FROM AUTHOR]

Massively open online courses (MOOCs) offer learners opportunities for self‐improvement and knowledge development. Linking cognitive processes and learning outcomes is helpful in supporting students learning with the help of MOOCs. Based on the cognitive presence of the Community of Inquiry framework, this study quantified the effect of cognitive presence on students' learning performance by analysing more than 400,000 posts of 13 MOOCs. First, the study built a highly predictive classification model using a machine learning algorithm to automatically identify the phases of cognitive presence in MOOC forum posts. Subsequently, multilevel modelling was used to analyse the influence of learners' cognitive presence on their learning performance. The results showed that different phases of cognitive presence influenced students' learning differently. The findings help us understand the phases and depth of students' cognitive presence and can be used as the basis for MOOCs to automatically provide appropriate cognitive feedback and support to students. Practitioner notesWhat is already known about this topic Massively open online courses (MOOCs) offer learners opportunities for self‐improvement and knowledge development.Linking cognitive processes and learning outcomes is helpful in supporting student learning in MOOCs.Little is known so far about the role of cognitive presence in influencing students' learning performance.What this paper adds This study built a highly predictive classification model using a machine learning algorithm to automatically identify the phases of cognitive presence in MOOC forum posts.Multilevel modelling was used to quantify the influence of learners' cognitive presence on their learning performance.The results showed that different phases of cognitive presence influenced students' learning differently.Implications for practice and/or policy This study built an effective, high‐performance machine learning model with easy feature extraction, which could be used to automatically identify students' cognitive processes in MOOC forums and beyond.The combination of cognitive presence and learning performance helps us understand the phases and depth of students' cognitive presence.The findings can be used as the basis for MOOCs to automatically provide appropriate cognitive feedback and support to students. [ABSTRACT FROM AUTHOR]

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

With the widespread use of learning analytics (LA), ethical concerns about fairness have been raised. Research shows that LA models may be biased against students of certain demographic subgroups. Although fairness has gained significant attention in the broader machine learning (ML) community in the last decade, it is only recently that attention has been paid to fairness in LA. Furthermore, the decision on which unfairness mitigation algorithm or metric to use in a particular context remains largely unknown. On this premise, we performed a comparative evaluation of some selected unfairness mitigation algorithms regarded in the fair ML community to have shown promising results. Using a 3‐year program dropout data from an Australian university, we comparatively evaluated how the unfairness mitigation algorithms contribute to ethical LA by testing for some hypotheses across fairness and performance metrics. Interestingly, our results show how data bias does not always necessarily result in predictive bias. Perhaps not surprisingly, our test for fairness‐utility tradeoff shows how ensuring fairness does not always lead to drop in utility. Indeed, our results show that ensuring fairness might lead to enhanced utility under specific circumstances. Our findings may to some extent, guide fairness algorithm and metric selection for a given context. Practitioner notesWhat is already known about this topic LA is increasingly being used to leverage actionable insights about students and drive student success.LA models have been found to make discriminatory decisions against certain student demographic subgroups—therefore, raising ethical concerns.Fairness in education is nascent. Only a few works have examined fairness in LA and consequently followed up with ensuring fair LA models.What this paper adds A juxtaposition of unfairness mitigation algorithms across the entire LA pipeline showing how they compare and how each of them contributes to fair LA.Ensuring ethical LA does not always lead to a dip in performance. Sometimes, it actually improves performance as well.Fairness in LA has only focused on some form of outcome equality, however equality of outcome may be possible only when the playing field is levelled.Implications for practice and/or policy Based on desired notion of fairness and which segment of the LA pipeline is accessible, a fairness‐minded decision maker may be able to decide which algorithm to use in order to achieve their ethical goals.LA practitioners can carefully aim for more ethical LA models without trading significant utility by selecting algorithms that find the right balance between the two objectives.Fairness enhancing technologies should be cautiously used as guides—not final decision makers. Human domain experts must be kept in the loop to handle the dynamics of transcending fair LA beyond equality to equitable LA. [ABSTRACT FROM AUTHOR]

This paper investigates the end-to-end human-Mars entry, powered-descent, and landing (EDL) guidance problem by developing an on-board learning-based optimal control method (L-OCM) to achieve the goal of precise and fuel-efficient planetary landing. First, the end-to-end EDL guidance problem is formulated as a multi-phase optimal control problem with hybrid dynamics and constraints. Then a customized alternating direction method of multipliers is applied to solve the end-to-end EDL guidance problem with varying initial states off-line. After that, the L-OCM is developed to generate real-time optimal guidance commands. To be specific, supported by the optimal control theory, the necessary conditions of optimality for optimal control of the entry phase and powered-descent phase are derived, respectively, which leads to two two-point-boundary-value-problems (TPBVPs). Then, critical parameters are identified to approximate the complete solutions of the TPBVPs. To find the implicit relationship between the initial states and these critical parameters, deep neural networks are constructed to learn the values of these critical parameters in real-time with training data obtained from the off-line solutions. Furthermore, when random disturbances are considered during the EDL process, the single stage L-OCM is extended to the multi-stage L-OCM to regenerate real-time optimal guidance commands. Finally, the proposed L-OCM is implemented in extensive simulation cases to verify the effectiveness and efficiency of the new method. • We consider the multi-phase guidance problem integrating the entry and powered descent phases. • The learning method only needs to learn identified critical parameters to reduce computation load. • The proposed approach can handle multi-type disturbances during the EDL mission. • The proposed method can achieve precise and fuel-efficient planetary landing in real-time. [ABSTRACT FROM AUTHOR]

Human Learning Optimization (HLO) is an efficient metaheuristic algorithm in which three learning operators, i.e., the random learning operator, the individual learning operator, and the social learning operator, are developed to search for optima by mimicking the learning behaviors of humans. In fact, people not only learn from global optimization but also learn from the best solution of other individuals in the real life, and the operators of Differential Evolution are updated based on the optima of other individuals. Inspired by these facts, this paper proposes two novel differential human learning optimization algorithms (DEHLOs), into which the Differential Evolution strategy is introduced to enhance the optimization ability of the algorithm. And the two optimization algorithms, based on improving the HLO from individual and population, are named DEHLO1 and DEHLO2, respectively. The multidimensional knapsack problems are adopted as benchmark problems to validate the performance of DEHLOs, and the results are compared with the standard HLO and Modified Binary Differential Evolution (MBDE) as well as other state-of-the-art metaheuristics. The experimental results demonstrate that the developed DEHLOs significantly outperform other algorithms and the DEHLO2 achieves the best overall performance on various problems. [ABSTRACT FROM AUTHOR]

The architecture of AND/OR fuzzy neural networks exhibits outstanding learning abilities and significant interpretation capabilities. However, AND/OR networks suffer from structure-related problems namely low efficiency and slow convergence of learning due to several reasons such as high dimensionality and gradient-based learning algorithms which lead to a visible computing overhead. In this paper, we present a two-phase fuzzy logic-oriented network design that is composed of AND/OR neurons. This design takes advantages of Randomized Neural Network (RNN) to achieve higher convergence while exhibiting good nonlinear approximation capabilities. A gradient–based learning algorithm is implemented in the second phase of the design to further reduce values of performance index. The quality of the proposed design and resulting architecture is quantified through the use of numeric data along with fuzzy sets (information granules). Experimental results meet the research's objectives and the proposed design methodology opens up new future directions for proceeding with more improvements. [ABSTRACT FROM AUTHOR]

Continual lifelong learning is an machine learning framework inspired by human learning, where learners are trained to continuously acquire new knowledge in a sequential manner. However, the non-stationary nature of streaming training data poses a significant challenge known as catastrophic forgetting, which refers to the rapid forgetting of previously learned knowledge when new tasks are introduced. While some approaches, such as experience replay (ER), have been proposed to mitigate this issue, their performance remains limited, particularly in the class-incremental scenario which is considered natural and highly challenging. In this paper, we present a novel algorithm, called adaptive-experience replay (AdaER), to address the challenge of continual lifelong learning. AdaER consists of two stages: memory replay and memory update. In the memory replay stage, AdaER introduces a contextually-cued memory recall (C-CMR) strategy, which selectively replays memories that are most conflicting with the current input data in terms of both data and task. Additionally, AdaER incorporates an entropy-balanced reservoir sampling (E-BRS) strategy to enhance the performance of the memory buffer by maximizing information entropy. To evaluate the effectiveness of AdaER, we conduct experiments on established supervised continual lifelong learning benchmarks, specifically focusing on class-incremental learning scenarios. The results demonstrate that AdaER outperforms existing continual lifelong learning baselines, highlighting its efficacy in mitigating catastrophic forgetting and improving learning performance. [ABSTRACT FROM AUTHOR]

In many engineered systems, optimization is used for decision making at time scales ranging from real-time operation to long-term planning. This process often involves solving similar optimization problems over and over again with slightly modified input parameters, often under tight latency requirements. We consider the problem of using the information available through this repeated solution process to learn important characteristics of the optimal solution as a function of the input parameters. Our proposed method is based on learning relevant sets of active constraints, from which the optimal solution can be obtained efficiently. Using active sets as features preserves information about the physics of the system, enables interpretable results, accounts for relevant safety constraints, and is easy to represent and encode. However, the total number of active sets is also very large, as it grows exponentially with system size. The key contribution of this paper is a streaming algorithm that learns the relevant active sets from training samples consisting of the input parameters and the corresponding optimal solution, without any restrictions on the problem type, problem structure or probability distribution of the input parameters. The algorithm comes with theoretical performance guarantees and is shown to converge fast for problem instances with a small number of relevant active sets. It can thus be used to establish simultaneously learn the relevant active sets and the practicability of the learning method. Through case studies in optimal power flow, supply chain planning, and shortest path routing, we demonstrate that often only a few active sets are relevant in practice, suggesting that active sets provide an appropriate level of abstraction for a learning algorithm to target. [ABSTRACT FROM AUTHOR]

Learning is the process of gaining knowledge and implementing this knowledge on behavior. The concept of learning is not strict to just human being, it expanded to include machine also. Now the machines can behave based on the gained knowledge learned from the environment. The learning process is evolving in both human and machine, to keep up with the technology in the world, the human learning evolved into micro-learning and the machine learning evolved to deep learning. In this paper, the evolution of learning is discussed as a formal survey accomplished with the foundation of machine learning and its evolved version of learning which is deep learning and micro-learning as a new learning technology can be implemented on human and machine learning. A procedural comparison is achieved to declare the purpose of this survey, also a related discussion integrates the aim of this study. Finally a concluded points are illustrated as outcome which summarized the practical evolution intervals of the machine learning different concepts. [ABSTRACT FROM AUTHOR]

The $r$ -parallel set of a set $A \subseteq {\mathbb {R}} ^{d}$ is the set of all points whose distance from $A$ is less than $r$. In this paper, we show that the surface area of an $r$ -parallel set in ${\mathbb {R}}^{d}$ with volume at most $V$ is upper-bounded by $e^{\Theta (d)}V/r$ , whereas its Gaussian surface area is upper-bounded by $\max (e^{\Theta (d)}, e^{\Theta (d)}/r)$. We also derive a reverse form of the Brunn-Minkowski inequality for $r$ -parallel sets, and as an aside a reverse entropy power inequality for Gaussian-smoothed random variables. We apply our results to two problems in theoretical machine learning: (1) bounding the computational complexity of learning $r$ -parallel sets under a Gaussian distribution; and (2) bounding the sample complexity of estimating robust risk, which is a notion of risk in the adversarial machine learning literature that is analogous to the Bayes risk in hypothesis testing. [ABSTRACT FROM AUTHOR]

In the last few years, the literature conferred a great interest in studying the feasibility of using memristive devices for computing. Memristive devices are important in structure, dynamics, as well as functionalities of artificial neural networks (ANNs) because of their resemblance to biological learning in synapses and neurons regarding switching characteristics of their resistance. Memristive architecture consists of a number of metastable switches (MSSs). Although the literature covered a variety of memristive applications for general purpose computations, the effect of low or high conductance of each MSS was unclear. This paper focuses on finding a potential criterion to calculate the conductance of each MMS rather than the whole conductance as reported in the literature. Anti-Hebbian and Hebbian (AHaH) learning rules are used to mimic the changes in memristance of the memristors. This research will concentrate on the effect of conductance on an individual MSS to simulate the nanotechnology devices of the memristors. A single synapse is presented by a couple of memristors to mimic its resistance switching. The learning circuit of artificial synapses could be used in many applications, such as image processing and neural networks, for pattern classification of synapses, represented by a map of the memeristors. These synapses are essential elements for data processing and information storage in both real and artificial neural systems. [ABSTRACT FROM AUTHOR]

Recently, most existing human pose estimation methods fuse multi-stage convolutional modules to learn a shared feature representation. In this paper, we propose a expectation–maximization (EM) mapping-based network to learn specific related body parts for human pose estimation, named EMposeNet. It maps specific feature of related parts from the original fully shared feature space. From the perspective of multi-task learning, we can regard the task of human pose estimation as a homogeneous multi-task learning. Sharing features among related tasks can result in a more compact model and better generalization ability. However, sharing features for those unrelated or weakly related tasks will deteriorate the estimation performance. Our proposed method aims at performing EM algorithm to learn the related body part, where the predicted keypoint heatmap is potentially more accurate and spatially more precise. We conduct extensive experiments on two benchmark datasets, including the MSCOCO keypoint detection dataset and the MPII human pose dataset, to empirically demonstrate the validity of the proposed method. The results on such two benchmark datasets show that the proposed approach achieves a competitive performance. [ABSTRACT FROM AUTHOR]

The attempt to understand the cerebellum has been dominated for years by supervised learning models. The central idea is that a learning algorithm modifies transmission strength at repeatedly co-active synapses, creating memories stored as finely calibrated synaptic weights. As a result, Purkinje cells, usually the de facto output cells of these models, acquire a modified response to input in a remembered pattern. This paper proposes an alternative model of pattern memory in which the function of a match is permissive, allowing but not driving output, and accordingly controlling the timing of output but not the rate of firing by Purkinje cells. Learning does not result in graded synaptic weights. There is no supervised learning algorithm or memory of individual patterns, which, like graded weights, are unnecessary to explain the evidence. Instead, patterns are classed as simply either known or not, at the level of input to a functional population of 100s of Purkinje cells (a microzone). The standard is strict. If only a handful of Purkinje cells receive a mismatch output of the whole circuit is blocked. Only if there is a full and accurate match are projection neurons in deep nuclei, which carry the output of most circuits, released from default inhibitory restraint. Purkinje cell firing at those times is a linear function of input rates. There is no effect of modification of synaptic transmission except to either allow or block output. [ABSTRACT FROM AUTHOR]

RGB-Infrared (RGB-IR) cross-modality person re-identification (re-ID) is attracting more and more attention due to requirements for 24-h scene surveillance. However, the high cost of labeling person identities of an RGB-IR dataset largely limits the scalability of supervised models in real-world scenarios. In this paper, we study the unsupervised RGB-IR person re-ID problem (or briefly uRGB-IR re-ID) in which no identity annotations are available in RGB-IR cross-modality datasets. Considering that intra-modality (i.e., RGB-RGB or IR-IR) re-ID is much easier than cross-modality re-ID and can provide shared knowledge for RGB-IR re-ID, we propose a two-stage method to solve the uRGB-IR re-ID, namely homogeneous-to-heterogeneous learning. In the first stage, the unsupervised self-learning method is conducted to learn the intra-modality feature representation and to generate the pseudo-labeled identities of person images separately for each modality. In the second stage, heterogeneous learning is used to learn a shared discriminative feature representation by distilling the knowledge from intra-modality pseudo-labels, to align two modalities via a modality-based consistent learning module, and finally to target modality-invariant learning via a pseudo-labeled positive instance selection module. With the use of homogeneous-to-heterogeneous learning, the proposed unsupervised framework greatly reduces the modality gap and thus learns a robust feature representation against RGB and infrared modalities, leading to promising accuracy. We also propose a novel cross-modality re-ranking approach that includes a self-modality search and a cycle-modality search to tailor the uRGB-IR re-ID. Unlike conventional re-ranking, the proposed re-ranking method takes a modality-based constraint into re-ranking and thus can select more reliable nearest neighbors, which greatly improves uRGB-IR re-ID. The experimental results demonstrate the superiority of our approach on the SYSU-MM01 and RegDB datasets. [ABSTRACT FROM AUTHOR]

When performing learning tasks in pervasive environments, the main challenge arises from the need of combining federated and continual settings. The former comes from the massive distribution of devices with privacy-regulated data. The latter is required by the low resources of the participating devices, which may retain data for short periods of time. In this paper, we propose a setup for learning with Echo State Networks (ESNs) in pervasive environments. Our proposal focuses on the use of Intrinsic Plasticity (IP), a gradient-based method for adapting the reservoir's non-linearity. First, we extend the objective function of IP to include the uncertainty arising from the distribution of the data over space and time. Then, we propose Federated Intrinsic Plasticity (FedIP), which is intended for client–server federated topologies with stationary data, and adapts the learning scheme provided by Federated Averaging (FedAvg) to include the learning rule of IP. Finally, we further extend this algorithm for learning to Federated Continual Intrinsic Plasticity (FedCLIP) to equip clients with CL strategies for dealing with continuous data streams. We evaluate our approach on an incremental setup built upon real-world datasets from human monitoring, where we tune the complexity of the scenario in terms of the distribution of the data over space and time. Results show that both our algorithms improve the representation capabilities and the performance of the ESN, while being robust to catastrophic forgetting. • Reservoir computing is very effective for continual and distributed learning. • Adding continual and federated learning to echo state networks. • Unsupervised federated and continual adaptation of reservoir. • Federated and continual learning for human action recognition and stress prediction. [ABSTRACT FROM AUTHOR]