Pub Date : 2024-03-31DOI: 10.1109/TAI.2024.3408130
Huan Li;Jun Fu;Tianyou Chai
To address the challenge of globally optimal control of path-constrained switched systems, a hybrid intelligent dynamic optimization method is proposed by combining the biobjective particle swarm optimization (PSO) method and a gradient descent method, which simultaneously obtains globally optimal switching instants and input and guarantees rigorous satisfaction of the path constraints over the continuous time horizon. First, the path constraint of switched systems is discretized into multiple point constraints, and then the right-hand side of the path constraint ( $leq 0$ ) is substituted with a negative value ( $leq-varepsilon$ ). Second, the single-objective constrained dynamic program of switched systems is transformed into a biobjective unconstrained dynamic program where each particle intelligently adjusts its objectives to detect the global optimum area satisfying the constraints, depending on its current position in the search space by the search mechanism of PSO. Third, the deterministic optimization method is deployed in the detected global optimum area to locate a feasible solution satisfying the Karush–Kuhn–Tucker (KKT) conditions to a specified tolerance of dynamic optimization of switched systems. Moreover, it is proved that the hybrid intelligent dynamic optimization method can obtain the optimal solution satisfying the first-order approximation KKT conditions within a finite number of iterations. Finally, the results of numerical simulations show the effectiveness of the presented method in terms of improving the solution accuracy and guaranteeing rigorous satisfaction of the path constraint.
{"title":"Hybrid Intelligent Optimization of Nonlinear Switched Systems With Guaranteed Feasibility","authors":"Huan Li;Jun Fu;Tianyou Chai","doi":"10.1109/TAI.2024.3408130","DOIUrl":"https://doi.org/10.1109/TAI.2024.3408130","url":null,"abstract":"To address the challenge of \u0000<italic>globally</i>\u0000 optimal control of path-constrained switched systems, a hybrid intelligent dynamic optimization method is proposed by combining the biobjective particle swarm optimization (PSO) method and a gradient descent method, which simultaneously obtains globally optimal switching instants and input and guarantees rigorous satisfaction of the path constraints over the continuous time horizon. First, the path constraint of switched systems is discretized into multiple point constraints, and then the right-hand side of the path constraint (\u0000<inline-formula><tex-math>$leq 0$</tex-math></inline-formula>\u0000) is substituted with a negative value (\u0000<inline-formula><tex-math>$leq-varepsilon$</tex-math></inline-formula>\u0000). Second, the single-objective constrained dynamic program of switched systems is transformed into a biobjective unconstrained dynamic program where each particle intelligently adjusts its objectives to detect the global optimum area satisfying the constraints, depending on its current position in the search space by the search mechanism of PSO. Third, the deterministic optimization method is deployed in the detected global optimum area to locate a feasible solution satisfying the Karush–Kuhn–Tucker (KKT) conditions to a specified tolerance of dynamic optimization of switched systems. Moreover, it is proved that the hybrid intelligent dynamic optimization method can obtain the optimal solution satisfying the first-order approximation KKT conditions within a finite number of iterations. Finally, the results of numerical simulations show the effectiveness of the presented method in terms of improving the solution accuracy and guaranteeing rigorous satisfaction of the path constraint.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"5 10","pages":"5244-5257"},"PeriodicalIF":0.0,"publicationDate":"2024-03-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142443089","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
An accurate estimation of future stock prices can help investors maximize their profits. The current advancements in the area of artificial intelligence (AI) have proven prevalent in the financial sector. Besides, stock market prediction is difficult owing to the considerable volatility and unpredictability induced by numerous factors. Recent approaches have considered fundamental, technical, or macroeconomic variables to find hidden complex patterns in financial data. At the macro level, there exists a spillover effect between stock pairs that can explain the variance present in the data and boost the prediction performance. To address this interconnectedness defined by intrasector stocks, we propose a hybrid relational approach to predict the future price of stocks in the American, Indian, and Korean economies. We collected market data of large-, mid-, and small-capitalization peer companies in the same industry as the target firm, considering them as relational features. To ensure efficient feature selection, we have utilized a data-driven approach, i.e., random forest feature permutation (RF2P), to remove noise and instability. A hybrid prediction module consisting of temporal convolution and linear model (TCLM) is proposed that considers irregularities and linear trend components of the financial data. We found that RF2P-TCLM gave the superior performance. To demonstrate the real-world applicability of our approach in terms of profitability, we created a trading method based on the predicted results. This technique generates a higher profit than the existing approaches.
{"title":"A Hybrid Relational Approach Toward Stock Price Prediction and Profitability","authors":"Manali Patel;Krupa Jariwala;Chiranjoy Chattopadhyay","doi":"10.1109/TAI.2024.3408129","DOIUrl":"https://doi.org/10.1109/TAI.2024.3408129","url":null,"abstract":"An accurate estimation of future stock prices can help investors maximize their profits. The current advancements in the area of artificial intelligence (AI) have proven prevalent in the financial sector. Besides, stock market prediction is difficult owing to the considerable volatility and unpredictability induced by numerous factors. Recent approaches have considered fundamental, technical, or macroeconomic variables to find hidden complex patterns in financial data. At the macro level, there exists a spillover effect between stock pairs that can explain the variance present in the data and boost the prediction performance. To address this interconnectedness defined by intrasector stocks, we propose a hybrid relational approach to predict the future price of stocks in the American, Indian, and Korean economies. We collected market data of large-, mid-, and small-capitalization peer companies in the same industry as the target firm, considering them as relational features. To ensure efficient feature selection, we have utilized a data-driven approach, i.e., random forest feature permutation (RF2P), to remove noise and instability. A hybrid prediction module consisting of temporal convolution and linear model (TCLM) is proposed that considers irregularities and linear trend components of the financial data. We found that RF2P-TCLM gave the superior performance. To demonstrate the real-world applicability of our approach in terms of profitability, we created a trading method based on the predicted results. This technique generates a higher profit than the existing approaches.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"5 11","pages":"5844-5854"},"PeriodicalIF":0.0,"publicationDate":"2024-03-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142600169","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-30DOI: 10.1109/TAI.2024.3404913
G. S. Asha Rani;P. S. Lal Priya
An increased demand in electric vehicle (EV) charging facilities has necessitated intelligent energy management systems (EMSs), to control and monitor the available energy sources in these charging stations. The goal is to create a charging schedule for EVs that minimizes the operating cost of the charging station while ensuring all connected EV's charging demands. Model predictive control (MPC) has been widely used for EMS. The challenge with MPC is that a precise representation of the underlying physical system's dynamics is essential. In this study, machine learning methods are combined with conventional MPC to build a data-driven MPC (DMPC) which can adapt to the changes in the system's behavior over time. As new data become available, the data-driven model can be updated and the MPC algorithm can be reoptimized to reflect the current behavior of the system. Ensemble learning is an effective machine learning technique that increases the effectiveness and accuracy of decision making by utilizing the combined knowledge of several models. Out of the several methods available for implementing ensemble learning, adaptive random forest (ARF) algorithm with affine functions and convex optimization is selected. The results show comparable performance of DMPC with respect to MPC implemented on a well-established mathematical model of the system.
{"title":"Data-Driven Model Predictive Control for Hybrid Charging Stations Using Ensemble Learning","authors":"G. S. Asha Rani;P. S. Lal Priya","doi":"10.1109/TAI.2024.3404913","DOIUrl":"https://doi.org/10.1109/TAI.2024.3404913","url":null,"abstract":"An increased demand in electric vehicle (EV) charging facilities has necessitated intelligent energy management systems (EMSs), to control and monitor the available energy sources in these charging stations. The goal is to create a charging schedule for EVs that minimizes the operating cost of the charging station while ensuring all connected EV's charging demands. Model predictive control (MPC) has been widely used for EMS. The challenge with MPC is that a precise representation of the underlying physical system's dynamics is essential. In this study, machine learning methods are combined with conventional MPC to build a data-driven MPC (DMPC) which can adapt to the changes in the system's behavior over time. As new data become available, the data-driven model can be updated and the MPC algorithm can be reoptimized to reflect the current behavior of the system. Ensemble learning is an effective machine learning technique that increases the effectiveness and accuracy of decision making by utilizing the combined knowledge of several models. Out of the several methods available for implementing ensemble learning, adaptive random forest (ARF) algorithm with affine functions and convex optimization is selected. The results show comparable performance of DMPC with respect to MPC implemented on a well-established mathematical model of the system.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"5 10","pages":"5304-5313"},"PeriodicalIF":0.0,"publicationDate":"2024-03-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142443110","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-29DOI: 10.1109/TAI.2024.3406289
Xinliang Zhou;Chenyu Liu;Ruizhi Yang;Liangwei Zhang;Liming Zhai;Ziyu Jia;Yang Liu
Epilepsy is a life-threatening and challenging neurological disorder, and applying an electroencephalogram (EEG) is a commonly used clinical approach for its detection. Neuropsychological research indicates that epilepsy seizures are highly associated with distinct ranges of temporal EEG patterns. Although previous attempts to automatically detect epilepsy have achieved high classification performance, one crucial challenge still remains: how to effectively learn the robust global-local representation associated with epilepsy in the signals? To address the above challenge, we propose global-local neural epilepsy detection network (GlepNet), a novel architecture for automatic EEG epilepsy detection. We interleave the temporal convolution model together with the multihead attention mechanism within the GlepNet's encoder blocks to jointly capture the interlaced epilepsy seizure local and global features in EEG signals. Meanwhile, the interpretable method, gradient-weighted class activation mapping (Grad-CAM), is applied to visually confirm that the GlepNet acquires the ability to accord significant weight to EEG segments containing epileptiform abnormalities such as spike-wave complexes. Specifically, the Grad-CAM heatmaps are generated by backpropagating the gradients from the encoder blocks to highlight the epilepsy seizure-related parts. Extensive experiments show the superiority of the GlepNet over state-of-the-art methods on multiple EEG epilepsy datasets. The code will soon be open-sourced on GitHub.
{"title":"Learning Robust Global-Local Representation From EEG for Neural Epilepsy Detection","authors":"Xinliang Zhou;Chenyu Liu;Ruizhi Yang;Liangwei Zhang;Liming Zhai;Ziyu Jia;Yang Liu","doi":"10.1109/TAI.2024.3406289","DOIUrl":"https://doi.org/10.1109/TAI.2024.3406289","url":null,"abstract":"Epilepsy is a life-threatening and challenging neurological disorder, and applying an electroencephalogram (EEG) is a commonly used clinical approach for its detection. Neuropsychological research indicates that epilepsy seizures are highly associated with distinct ranges of temporal EEG patterns. Although previous attempts to automatically detect epilepsy have achieved high classification performance, one crucial challenge still remains: how to effectively learn the robust global-local representation associated with epilepsy in the signals? To address the above challenge, we propose global-local neural epilepsy detection network (GlepNet), a novel architecture for automatic EEG epilepsy detection. We interleave the temporal convolution model together with the multihead attention mechanism within the GlepNet's encoder blocks to jointly capture the interlaced epilepsy seizure local and global features in EEG signals. Meanwhile, the interpretable method, gradient-weighted class activation mapping (Grad-CAM), is applied to visually confirm that the GlepNet acquires the ability to accord significant weight to EEG segments containing epileptiform abnormalities such as spike-wave complexes. Specifically, the Grad-CAM heatmaps are generated by backpropagating the gradients from the encoder blocks to highlight the epilepsy seizure-related parts. Extensive experiments show the superiority of the GlepNet over state-of-the-art methods on multiple EEG epilepsy datasets. The code will soon be open-sourced on GitHub.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"5 11","pages":"5720-5732"},"PeriodicalIF":0.0,"publicationDate":"2024-03-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142600099","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-27DOI: 10.1109/TAI.2024.3405405
Zixiang Wei;Yiting Wang;Lichao Sun;Athanasios V. Vasilakos;Lin Wang
Low-light images often suffer from limited visibility and multiple types of degradation, rendering low-light image enhancement (LIE) a nontrivial task. Some endeavors have been made to enhance low-light images using convolutional neural networks (CNNs). However, they have low efficiency in learning the structural information and diverse illumination levels at the local regions of an image. Consequently, the enhanced results are affected by unexpected artifacts, such as unbalanced exposure, blur, and color bias. This article proposes a novel framework, called ClassLIE, that combines the potential of CNNs and transformers. It classifies and adaptively learns the structural and illumination information from the low-light images in a holistic and regional manner, thus showing better enhancement performance. Our framework first employs a structure and illumination classification (SIC) module to learn the degradation information adaptively. In SIC, we decompose an input image into an illumination map and a reflectance map. A class prediction block is then designed to classify the degradation information by calculating the structure similarity scores on the reflectance map and mean square error (MSE) on the illumination map. As such, each input image can be divided into patches with three enhancement difficulty levels. Then, a feature learning and fusion (FLF) module is proposed to adaptively learn the feature information with CNNs for different enhancement difficulty levels while learning the long-range dependencies for the patches in a holistic manner. Experiments on five benchmark datasets consistently show our ClassLIE achieves new state-of-the-art performance, with 25.74 peak signal-to-noise ratio (PSNR) and 0.92 structural similarity (SSIM) on the LOw-Light (LOL) dataset.
{"title":"ClassLIE: Structure- and Illumination-Adaptive Classification for Low-Light Image Enhancement","authors":"Zixiang Wei;Yiting Wang;Lichao Sun;Athanasios V. Vasilakos;Lin Wang","doi":"10.1109/TAI.2024.3405405","DOIUrl":"https://doi.org/10.1109/TAI.2024.3405405","url":null,"abstract":"Low-light images often suffer from limited visibility and multiple types of degradation, rendering low-light image enhancement (LIE) a nontrivial task. Some endeavors have been made to enhance low-light images using convolutional neural networks (CNNs). However, they have low efficiency in learning the structural information and diverse illumination levels at the local regions of an image. Consequently, the enhanced results are affected by unexpected artifacts, such as unbalanced exposure, blur, and color bias. This article proposes a novel framework, called ClassLIE, that combines the potential of CNNs and transformers. It classifies and adaptively learns the structural and illumination information from the low-light images in a holistic and regional manner, thus showing better enhancement performance. Our framework first employs a structure and illumination classification (SIC) module to learn the degradation information adaptively. In SIC, we decompose an input image into an illumination map and a reflectance map. A class prediction block is then designed to classify the degradation information by calculating the structure similarity scores on the reflectance map and mean square error (MSE) on the illumination map. As such, each input image can be divided into patches with three enhancement difficulty levels. Then, a feature learning and fusion (FLF) module is proposed to adaptively learn the feature information with CNNs for different enhancement difficulty levels while learning the long-range dependencies for the patches in a holistic manner. Experiments on five benchmark datasets consistently show our ClassLIE achieves new state-of-the-art performance, with 25.74 peak signal-to-noise ratio (PSNR) and 0.92 structural similarity (SSIM) on the LOw-Light (LOL) dataset.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"5 9","pages":"4765-4775"},"PeriodicalIF":0.0,"publicationDate":"2024-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142169750","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-27DOI: 10.1109/TAI.2024.3405468
Akash C. Babu;Jeevanand Seshadrinath
This work introduces a novel online signal processing and machine learning (ML) framework designed for the incipient diagnosis of stator interturn faults (SITF) in three-phase squirrel cage induction motors. Addressing the critical need for incipient fault detection to prevent severe motor damage, the framework focuses on motor speed estimation, incipient fault detection, fault severity estimation, and faulty phase identification using only stator currents. A distinctive contribution lies in the proposed interacting multiple model (IMM) framework that leverages carefully selected motor current signatures as features, offering a comprehensive strategy for stator fault diagnosis not explored previously. The article pioneers the use of the selected harmonics with ML models to estimate a fault severity indicator, which is developed based on insights from the motor's physics of failure. Experimental validation showcases the fault indicator's effectiveness under diverse operating conditions, demonstrating its utility in fault severity assessment. Suitable standalone ML model is selected, or an ensemble is constructed from a pool of ML models at each stage of the IMM framework. Further, a feature relevance analysis is also performed to garner insights into the contributions of each handpicked feature in predicting the fault indicator.
本研究介绍了一种新型在线信号处理和机器学习（ML）框架，该框架专为三相鼠笼式感应电机定子匝间故障（SITF）的初期诊断而设计。为了满足初期故障检测的关键需求，防止严重的电机损坏，该框架重点关注电机速度估计、初期故障检测、故障严重性估计以及仅使用定子电流的故障相位识别。其独特之处在于提出了交互式多模型 (IMM) 框架，该框架利用精心选择的电机电流特征，为定子故障诊断提供了一种前所未有的综合策略。文章开创性地将选定的谐波与多模型（ML）模型结合使用，以估算故障严重性指标，该指标是基于对电机故障物理原理的深入了解而开发的。实验验证展示了故障指标在不同运行条件下的有效性，证明了其在故障严重性评估中的实用性。在 IMM 框架的每个阶段，都会选择合适的独立 ML 模型，或从 ML 模型池中构建一个集合。此外，还进行了特征相关性分析，以深入了解每个精选特征在预测故障指标方面的贡献。
{"title":"Interacting Multiple Model Framework for Incipient Diagnosis of Interturn Faults in Induction Motors","authors":"Akash C. Babu;Jeevanand Seshadrinath","doi":"10.1109/TAI.2024.3405468","DOIUrl":"https://doi.org/10.1109/TAI.2024.3405468","url":null,"abstract":"This work introduces a novel online signal processing and machine learning (ML) framework designed for the incipient diagnosis of stator interturn faults (SITF) in three-phase squirrel cage induction motors. Addressing the critical need for incipient fault detection to prevent severe motor damage, the framework focuses on motor speed estimation, incipient fault detection, fault severity estimation, and faulty phase identification using only stator currents. A distinctive contribution lies in the proposed interacting multiple model (IMM) framework that leverages carefully selected motor current signatures as features, offering a comprehensive strategy for stator fault diagnosis not explored previously. The article pioneers the use of the selected harmonics with ML models to estimate a fault severity indicator, which is developed based on insights from the motor's physics of failure. Experimental validation showcases the fault indicator's effectiveness under diverse operating conditions, demonstrating its utility in fault severity assessment. Suitable standalone ML model is selected, or an ensemble is constructed from a pool of ML models at each stage of the IMM framework. Further, a feature relevance analysis is also performed to garner insights into the contributions of each handpicked feature in predicting the fault indicator.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"5 10","pages":"5120-5129"},"PeriodicalIF":0.0,"publicationDate":"2024-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142443086","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-27DOI: 10.1109/TAI.2024.3382248
Khurram Khan;Atiq ur Rehman;Adnan Khan;Syed Rameez Naqvi;Samir Brahim Belhaouari;Amine Bermak
This work proposes a novel split and kernel-merge clustering (S-KMC), a nonparametric clustering algorithm that combines the strengths of hierarchical clustering, partitional clustering, and density-based clustering. It consists of two main phases: splitting and merging. In the splitting phase, a ranking-based operator is used to divide the data into optimal subclusters. In the merging phase, a kernel function estimates the density of these subclusters after projecting them onto a straight line passing through their centers, facilitating the merging operation. S-KMC is fully nonparametric, eliminating the need for prior information about the data. It effectively handles 1) shape diversity, 2) density variability, 3) high dimensionality, 4) outliers, and 5) missing values. The algorithm offers easily tunable hyperparameters, enhancing its applicability to complex problems and robustness against data anomalies. Experimental analysis on 21 benchmark datasets demonstrates the improved performance of S-KMC in terms of cluster accuracy, handling high-dimensional data, and managing data anomalies and outliers. Comprehensive comparisons with state-of-the-art techniques further validate the superior or comparable performance of the proposed S-KMC algorithm.
{"title":"A Nonparametric Split and Kernel-Merge Clustering Algorithm","authors":"Khurram Khan;Atiq ur Rehman;Adnan Khan;Syed Rameez Naqvi;Samir Brahim Belhaouari;Amine Bermak","doi":"10.1109/TAI.2024.3382248","DOIUrl":"https://doi.org/10.1109/TAI.2024.3382248","url":null,"abstract":"This work proposes a novel split and kernel-merge clustering (S-KMC), a nonparametric clustering algorithm that combines the strengths of hierarchical clustering, partitional clustering, and density-based clustering. It consists of two main phases: splitting and merging. In the splitting phase, a ranking-based operator is used to divide the data into optimal subclusters. In the merging phase, a kernel function estimates the density of these subclusters after projecting them onto a straight line passing through their centers, facilitating the merging operation. S-KMC is fully nonparametric, eliminating the need for prior information about the data. It effectively handles 1) shape diversity, 2) density variability, 3) high dimensionality, 4) outliers, and 5) missing values. The algorithm offers easily tunable hyperparameters, enhancing its applicability to complex problems and robustness against data anomalies. Experimental analysis on 21 benchmark datasets demonstrates the improved performance of S-KMC in terms of cluster accuracy, handling high-dimensional data, and managing data anomalies and outliers. Comprehensive comparisons with state-of-the-art techniques further validate the superior or comparable performance of the proposed S-KMC algorithm.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"5 9","pages":"4443-4457"},"PeriodicalIF":0.0,"publicationDate":"2024-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142164997","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-27DOI: 10.1109/TAI.2024.3405404
Shilin Zhang;Chenlin Yi
This work tackles the significant challenges inherent in online continual learning (OCL), a domain characterized by its handling of numerous tasks over extended periods. OCL is designed to adapt evolving data distributions and previously unseen classes through a single-pass analysis of a data stream, mirroring the dynamic nature of real-world applications. Despite its promising potential, existing OCL methodologies often suffer from catastrophic forgetting (CF) when confronted with a large array of tasks, compounded by substantial computational demands that limit their practical utility. At the heart of our proposed solution is the adoption of a kernel density estimation (KDE) learning framework, aimed at resolving the task prediction (TP) dilemma and ensuring the separability of all tasks. This is achieved through the incorporation of a linear projection head and a probability density function (PDF) for each task, while a shared backbone is maintained across tasks to provide raw feature representation. During the inference phase, we leverage an ensemble of PDFs, which utilizes a self-reporting mechanism based on maximum PDF values to identify the most appropriate model for classifying incoming instances. This strategy ensures that samples with identical labels are cohesively grouped within higher density PDF regions, effectively segregating dissimilar instances across the feature space of different tasks. Extensive experimental validation across diverse OCL datasets has underscored our framework's efficacy, showcasing remarkable performance enhancements and significant gains over existing methodologies, all achieved with minimal time-space overhead. Our approach introduces a scalable and efficient paradigm for OCL, addressing both the challenge of CF and computational efficiency, thereby extending the applicability of OCL to more realistic and demanding scenarios.
在线持续学习（OCL）是一个以长时间处理大量任务为特点的领域，这项工作旨在应对在线持续学习中固有的重大挑战。OCL 旨在通过对数据流的一次分析，适应不断变化的数据分布和以前未见过的类别，从而反映真实世界应用的动态性质。尽管 OCL 潜力巨大，但现有的 OCL 方法在面对大量任务时往往会出现灾难性遗忘 (CF)，再加上大量的计算需求限制了其实用性。我们提出的解决方案的核心是采用核密度估计（KDE）学习框架，旨在解决任务预测（TP）难题并确保所有任务的可分离性。这是通过为每项任务加入线性投影头和概率密度函数（PDF）来实现的，同时在各项任务中保持共享主干，以提供原始特征表示。在推理阶段，我们利用 PDF 集合，利用基于最大 PDF 值的自我报告机制来确定最适合对输入实例进行分类的模型。这一策略可确保将具有相同标签的样本凝聚到密度更高的 PDF 区域中，从而有效隔离不同任务特征空间中的异类实例。在不同的 OCL 数据集上进行的广泛实验验证证明了我们框架的功效，与现有方法相比，我们的框架显著提高了性能，并取得了巨大的收益，而所有这些都是在最小的时间空间开销下实现的。我们的方法为 OCL 引入了一种可扩展的高效范式，同时解决了 CF 和计算效率方面的挑战，从而将 OCL 的适用性扩展到了更现实、要求更高的场景中。
{"title":"Online Continual Learning Benefits From Large Number of Task Splits","authors":"Shilin Zhang;Chenlin Yi","doi":"10.1109/TAI.2024.3405404","DOIUrl":"https://doi.org/10.1109/TAI.2024.3405404","url":null,"abstract":"This work tackles the significant challenges inherent in online continual learning (OCL), a domain characterized by its handling of numerous tasks over extended periods. OCL is designed to adapt evolving data distributions and previously unseen classes through a single-pass analysis of a data stream, mirroring the dynamic nature of real-world applications. Despite its promising potential, existing OCL methodologies often suffer from catastrophic forgetting (CF) when confronted with a large array of tasks, compounded by substantial computational demands that limit their practical utility. At the heart of our proposed solution is the adoption of a kernel density estimation (KDE) learning framework, aimed at resolving the task prediction (TP) dilemma and ensuring the separability of all tasks. This is achieved through the incorporation of a linear projection head and a probability density function (PDF) for each task, while a shared backbone is maintained across tasks to provide raw feature representation. During the inference phase, we leverage an ensemble of PDFs, which utilizes a self-reporting mechanism based on maximum PDF values to identify the most appropriate model for classifying incoming instances. This strategy ensures that samples with identical labels are cohesively grouped within higher density PDF regions, effectively segregating dissimilar instances across the feature space of different tasks. Extensive experimental validation across diverse OCL datasets has underscored our framework's efficacy, showcasing remarkable performance enhancements and significant gains over existing methodologies, all achieved with minimal time-space overhead. Our approach introduces a scalable and efficient paradigm for OCL, addressing both the challenge of CF and computational efficiency, thereby extending the applicability of OCL to more realistic and demanding scenarios.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"5 11","pages":"5746-5759"},"PeriodicalIF":0.0,"publicationDate":"2024-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142600430","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
The accurate and rapid transfer of complex urban spatiotemporal data is crucial for urban computing tasks such as urban planning and public transportation deployment for smart-city applications. Existing works consider auxiliary data or propose end-to-end models to process complex spatiotemporal information into more complex deep features. However, the latter is incapable of decoupling spatiotemporal knowledge, which means these end-to-end models lack modularity and substitutability. A general modular framework that can automatically capture simple representations of complex spatiotemporal information is required. In this article, we thus propose a universal framework for the transfer of spatiotemporal knowledge based on a radial basis function (RBF). We termed this approach spatial–temporal RBF transfer framework (STRBF-TF). The proposed STRBF-TF generates simple RBF representations of spatiotemporal flow distribution with an RBF transfer block and also leverages a channel attention mechanism. Moreover, we propose two RBF kernel initializers suitable for the source and the target domains, respectively. The framework retains important spatiotemporal knowledge in simple representations for the reconfiguration of spatiotemporal feature distribution for fast and accurate transfer. We conducted cross-domain learning experiments on a large real-world telecom dataset. The results demonstrate the efficiency and accuracy of the proposed approach, as well as its suitability for real-world applications.
{"title":"Universal Transfer Framework for Urban Spatiotemporal Knowledge Based on Radial Basis Function","authors":"Sheng-Min Chiu;Yow-Shin Liou;Yi-Chung Chen;Chiang Lee;Rong-Kang Shang;Tzu-Yin Chang;Roger Zimmermann","doi":"10.1109/TAI.2024.3382267","DOIUrl":"https://doi.org/10.1109/TAI.2024.3382267","url":null,"abstract":"The accurate and rapid transfer of complex urban spatiotemporal data is crucial for urban computing tasks such as urban planning and public transportation deployment for smart-city applications. Existing works consider auxiliary data or propose end-to-end models to process complex spatiotemporal information into more complex deep features. However, the latter is incapable of decoupling spatiotemporal knowledge, which means these end-to-end models lack modularity and substitutability. A general modular framework that can automatically capture simple representations of complex spatiotemporal information is required. In this article, we thus propose a universal framework for the transfer of spatiotemporal knowledge based on a radial basis function (RBF). We termed this approach spatial–temporal RBF transfer framework (STRBF-TF). The proposed STRBF-TF generates simple RBF representations of spatiotemporal flow distribution with an RBF transfer block and also leverages a channel attention mechanism. Moreover, we propose two RBF kernel initializers suitable for the source and the target domains, respectively. The framework retains important spatiotemporal knowledge in simple representations for the reconfiguration of spatiotemporal feature distribution for fast and accurate transfer. We conducted cross-domain learning experiments on a large real-world telecom dataset. The results demonstrate the efficiency and accuracy of the proposed approach, as well as its suitability for real-world applications.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"5 9","pages":"4458-4469"},"PeriodicalIF":0.0,"publicationDate":"2024-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142164999","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Practical object detection systems are highly desired to be open-ended for learning on frequently evolved datasets. Moreover, learning with little supervision further adds flexibility for real-world applications such as autonomous driving and robotics, where large-scale datasets could be prohibitive or expensive to obtain. However, continual adaption with small training examples often results in catastrophic forgetting and dramatic overfitting. To address such issues, a compositional learning system is proposed to enable effective incremental object detection from nonstationary and few-shot data streams. First of all, a novel bilateral–head framework is proposed to decouple the representation learning of base (pretrained) and novel (few-shot) classes into separate embedding spaces, which takes care of novel concept integration and base knowledge retention simultaneously. Moreover, to enhance learning stability, a robust parameter updating rule, i.e., recall and progress mechanism, is carried out to constrain the optimization trajectory of sequential model adaption. Beyond that, to enforce intertask class discrimination with little memory burden, we present a between-class regularization method that expands the decision space of few-shot classes for constructing unbiased feature representation. Final, we deeply investigate the incomplete annotation issue considering the realistic scenario of incremental few-shot object detection (iFSOD) and propose a semisupervised object labeling mechanism to accurately recover the missing annotations for previously encountered classes, which further enhances the robustness of the target detector to counteract catastrophic forgetting. Extensive experiments conducted on both Pascal visual object classes dataset (VOC) and microsoft common objects in context dataset (MS-COCO) datasets demonstrate the effectiveness of our method.
{"title":"Bilateral-Head Region-Based Convolutional Neural Networks: A Unified Approach for Incremental Few-Shot Object Detection","authors":"Yiting Li;Haiyue Zhu;Sichao Tian;Jun Ma;Cheng Xiang;Prahlad Vadakkepat","doi":"10.1109/TAI.2024.3381919","DOIUrl":"https://doi.org/10.1109/TAI.2024.3381919","url":null,"abstract":"Practical object detection systems are highly desired to be open-ended for learning on frequently evolved datasets. Moreover, learning with little supervision further adds flexibility for real-world applications such as autonomous driving and robotics, where large-scale datasets could be prohibitive or expensive to obtain. However, continual adaption with small training examples often results in catastrophic forgetting and dramatic overfitting. To address such issues, a compositional learning system is proposed to enable effective incremental object detection from nonstationary and few-shot data streams. First of all, a novel bilateral–head framework is proposed to decouple the representation learning of base (pretrained) and novel (few-shot) classes into separate embedding spaces, which takes care of novel concept integration and base knowledge retention simultaneously. Moreover, to enhance learning stability, a robust parameter updating rule, i.e., recall and progress mechanism, is carried out to constrain the optimization trajectory of sequential model adaption. Beyond that, to enforce intertask class discrimination with little memory burden, we present a between-class regularization method that expands the decision space of few-shot classes for constructing unbiased feature representation. Final, we deeply investigate the incomplete annotation issue considering the realistic scenario of incremental few-shot object detection (iFSOD) and propose a semisupervised object labeling mechanism to accurately recover the missing annotations for previously encountered classes, which further enhances the robustness of the target detector to counteract catastrophic forgetting. Extensive experiments conducted on both Pascal visual object classes dataset (VOC) and microsoft common objects in context dataset (MS-COCO) datasets demonstrate the effectiveness of our method.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"5 9","pages":"4376-4390"},"PeriodicalIF":0.0,"publicationDate":"2024-03-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142165010","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：