Offline reinforcement learning (RL) can learn policies from pre-collected offline datasets without interacting with the environment, but it suffers from the issue of out-of-distribution (OOD). Recent methods use the generative adversarial paradigm to learn policies, but easily fail to handle the conflict of fooling the discriminator and maximizing expected returns. In this paper, we propose a novel offline RL method named Distribution-Matching Generator-based Diffusion Policies (DMGDP). A distribution matching-based policy learning method is first developed, where the diffusion serves as the policy generator, to handle the conflict of fooling the discriminator and maximizing expected returns. Furthermore, a policy confidence mechanism based on discriminator regularization is designed to prevent the agent from taking OOD actions, with the aim of robust generative adversarial learning. We conducted extensive experiments on the D4RL benchmarks, and the results demonstrate that DMGDP outperforms state-of-the-art methods.

Retrieval-Augmented Generation (RAG) improves large language models by retrieving external knowledge, often truncated into smaller chunks due to the input context window, which leads to information loss, resulting in response hallucinations and broken reasoning chains. Moreover, traditional RAG retrieves unstructured knowledge, introducing irrelevant details that hinder accurate reasoning. To address these issues, we propose TAdaRAG, a novel RAG framework for on-the-fly task-adaptive knowledge graph construction from external sources. Specifically, we design an intent-driven routing mechanism to a domain-specific extraction template, followed by supervised fine-tuning and a reinforcement learning-based implicit extraction mechanism, ensuring concise, coherent, and non-redundant knowledge integration. Evaluations on six public benchmarks and a real-world business benchmark (NowNewsQA) across three backbone models demonstrate that TAdaRAG outperforms existing methods across diverse domains and long-text tasks, highlighting its strong generalization and practical effectiveness.

Jihao Zhao, Ding Chen, Zhaoxin Fan, Kerun Xu, Mengting Hu, Bo Tang, Feiyu Xiong, Zhiyu li. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Multi-User MIMO (MU-MIMO) systems in consumer devices suffer from imperfect or delayed Channel State Information (CSI), limiting the effectiveness of conventional combiners. This paper introduces a two-stage learning framework that combines a Convolutional Neural Network-Gated Recurrent Unit (CNN–GRU) encoder for spatio-temporal CSI feature extraction with a Proximal Policy Optimization agent for adaptive, power-constrained combining. Simulations using 3GPP channel models with one-step CSI delay show that the proposed method enhances spectral efficiency by over 35% compared to Matched Filtering and by more than 10% over Minimum Mean Square Error, offering robust performance for interference-limited and dynamic environments.

Millimeter-wave (mmWave) communication systems face increasing susceptibility to advanced beam-stealing attacks, posing a significant physical layer security threat. This paper introduces a novel framework employing an advanced Deep Reinforcement Learning (DRL) agent for proactive and adaptive defense against these sophisticated attacks. A key innovation is leveraging Integrated Sensing and Communications (ISAC) capabilities for active, intelligent threat assessment. The DRL agent, built on a Proximal Policy Optimization (PPO) algorithm, dynamically controls ISAC probing actions to investigate suspicious activities. We introduce an intensive curriculum learning strategy that guarantees the agent experiences successful detection during training to overcome the complex exploration challenges inherent to such a security-critical task. Consequently, the agent learns a robust and adaptive policy that intelligently balances security and communication performance. Numerical results demonstrate that our framework achieves a mean attacker detection rate of 92.8% while maintaining an average user SINR of over 13 dB.

A classification rule based on the cumulant generating function of the training data, called the Cumulant Generating Function Classifier (CGFC), has been recently proposed, and has shown promising performance in terms of improved classification accuracy and robustness against noises. This paper first presents a new information-theoretical explanation of CGFC which indeed makes a classification by minimizing sample mutual information. The original CGFC is a type of global model, and a new variant, called Local-CGFC, is further introduced in this paper to achieve a local classification rule. Experimental studies on real-life datasets demonstrate the effectiveness of the proposed classifier and further illustrate its great potential for a number of real-world applications.

Safe reinforcement learning (RL) is developed to handle high-risk decision-making tasks, such as autonomous driving (AD), by constraining expected safety violation costs as a training objective. However, existing safe RL methods only consider the long-term objective but ignore the short-term state safety of exploration in the training process. In addition, it is difficult to achieve a balance between cost and return expectations, leading to deterioration of learning performance. Unlike these methods, we propose a novel algorithm named long-and short-term constraints (LSTCs) for safe RL. The short-term constraint is proposed to enhance the short-term state safety that the vehicle explores, while the long-term constraint enhances the overall safety of the vehicle throughout the decision-making process, both of which are jointly used to enhance vehicle safety in the training process. Furthermore, we develop a safe RL method with dual-constraint optimization based on the Lagrange multiplier to optimize the training process for end-to-end AD, balancing the cost and return expectations. Comprehensive experiments were conducted on the MetaDrive simulator. The experimental results demonstrate that the success rate increases by 13% and the episode cost decreases by 0.26 compared to the best results of the comparative methods, showing that the proposed method has better safety in continuous control tasks and exhibits a higher exploration performance in long-distance decision-making tasks compared to SOTA methods.

Xun Liang, Simin Niu, Zhiyu Li, Sensen Zhang, Hanyu Wang, Feiyu Xiong, Zhaoxin Fan, Bo Tang, Jihao Zhao, Jiawei Yang, Shichao Song, Mengwei Wang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

The evaluation of large language models (LLMs) has traditionally relied on static benchmarks, a paradigm that poses two major limitations: (1) predefined test sets lack adaptability to diverse application domains, and (2) standardized evaluation protocols often fail to capture fine-grained assessments of domain-specific knowledge and contextual reasoning abilities. To overcome these challenges, we propose GuessArena, an adaptive evaluation framework grounded in adversarial game-based interactions. Inspired by the interactive structure of the Guess Who I Am? game, our framework seamlessly integrates dynamic domain knowledge modeling with progressive reasoning assessment to improve evaluation fidelity. Empirical studies across five vertical domains-finance, healthcare, manufacturing, information technology, and education-demonstrate that GuessArena effectively distinguishes LLMs in terms of domain knowledge coverage and reasoning chain completeness. Compared to conventional benchmarks, our method provides substantial advantages in interpretability, scalability, and scenario adaptability.

Retrieval-Augmented Generation (RAG), while serving as a viable complement to large language models (LLMs), often overlooks the crucial aspect of text chunking within its pipeline. This paper initially introduces a dual-metric evaluation method, comprising Boundary Clarity and Chunk Stickiness, to enable the direct quantification of chunking quality. Leveraging this assessment method, we highlight the inherent limitations of traditional and semantic chunking in handling complex contextual nuances, thereby substantiating the necessity of integrating LLMs into chunking process. To address the inherent trade-off between computational efficiency and chunking precision in LLM-based approaches, we devise the granularity-aware Mixture-of-Chunkers (MoC) framework, which consists of a three-stage processing mechanism. Notably, our objective is to guide the chunker towards generating a structured list of chunking regular expressions, which are subsequently employed to extract chunks from the original text. Extensive experiments demonstrate that both our proposed metrics and the MoC framework effectively settle challenges of the chunking task, revealing the chunking kernel while enhancing the performance of the RAG system.

Retrieval-augmented citation generation (RACG) helps users trust the large language model output by retrieving evidence from reliable sources. However, most current RACG research focuses on single-language tasks, particularly in English, and overlooks the need for cross-lingual evidence retrieval and utilization in real-world applications. To address this issue, we introduce a plug-and-play Retrieval-Augmented Multilingual Citation Generation method (RAMCG) which uses a multilingual retriever to identify relevant evidence from a multilingual knowledge base. The evidence is then combined with the query and processed by a multilingual citation generator. The result is citations that are both accurate and comprehensive. Experiments show that RAMCG outperforms baseline methods in multilingual citation generation and is well-suited for practical use.

The continuous advancement of large language models (LLMs) has brought increasing attention to the critical issue of developing fair and reliable methods for evaluating their performance. Particularly, the emergence of cheating phenomena, such as test set leakage and prompt format overfitting, poses significant challenges to the reliable evaluation of LLMs. As evaluation frameworks commonly use Regular Expression (RegEx) for answer extraction, models may adjust their responses to fit formats easily handled by RegEx. Nevertheless, the key answer extraction module based on RegEx frequently suffers from extraction errors. Furthermore, recent studies proposing fine-tuned LLMs as judge models for automated evaluation face challenges in terms of generalization ability and fairness. This paper comprehensively analyzes the entire LLM evaluation chain and demonstrates that optimizing the key answer extraction module improves extraction accuracy and enhances evaluation reliability. Our findings suggest that improving the key answer extraction module can lead to higher judgment accuracy and improved evaluation efficiency compared to the judge models. To address these issues, we propose xFinder, a novel evaluator for answer extraction and matching in LLM evaluation. As part of this process, we create a specialized dataset, the Key Answer Finder (KAF) dataset, to ensure effective model training and evaluation. Generalization tests and real-world evaluations show that the smallest xFinder model, with only 500 million parameters, achieves an average extraction accuracy of 93.42\%. In contrast, RegEx accuracy in the best evaluation framework is 74.38\%. The final judgment accuracy of xFinder reaches 97.61\%, outperforming existing evaluation frameworks and judge models.

With the rapid development of Large Language Models (LLMs), aligning these models with human preferences and values is critical to ensuring ethical and safe applications. However, existing alignment techniques such as RLHF or DPO often require direct fine-tuning on LLMs with billions of parameters, resulting in substantial computational costs and inefficiencies. To address this, we propose Micro token-level Accept-Reject Aligning (MARA) approach designed to operate independently of the language models. MARA simplifies the alignment process by decomposing sentence-level preference learning into token-level binary classification, where a compact three-layer fully-connected network determines whether candidate tokens are “Accepted” or “Rejected” as part of the response. Extensive experiments across seven different LLMs and three open-source datasets show that MARA achieves significant improvements in alignment performance while reducing computational costs. The source code and implementation details are publicly available at https://github.com/IAAR-Shanghai/MARA, and the trained models are released at https://huggingface.co/IAAR-Shanghai/MARA_AGENTS.

Miao Li, Ming-Bin Chen, Bo Tang, ShengbinHou ShengbinHou, Pengyu Wang, Haiying Deng, Zhiyu Li, Feiyu Xiong, Keming Mao, Cheng Peng, Yi Luo. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Xun Liang, Shichao Song, Simin Niu, Zhiyu Li, Feiyu Xiong, Bo Tang, Yezhaohui Wang, Dawei He, Cheng Peng, Zhonghao Wang, Haiying Deng. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

The rapid advancement of 5G networks and the upcoming transition to 6G necessitate the use of the Open Radio Access Network (O-RAN) architecture to enable greater flexibility, interoperability, and innovation. This shift towards 6G and O-RAN requires the development of advanced simulation tools for testing, analyzing, and optimizing Radio Access Network (RAN) operations. This need becomes critical due to the complex dynamics of mobility management inherent in the 6G vision and next-generation networks. These networks anticipate advanced handover methods for mobile users, UAVs, IoT devices, and beyond. Addressing this gap, this paper introduces RANFusion: a robust RAN simulator specifically created to explore a variety of handover scenarios and to test and balance resources between users. This tool enables precise simulations for refining handover strategies within RAN and O-RAN environments, thereby ensuring optimal performance and reliability in these advanced network infrastructures.

Federated learning (FL) is a decentralized machine learning architecture, which leverages a large number of remote devices to learn a joint model with distributed training data. However, the system-heterogeneity is one major challenge in an FL network to achieve robust distributed learning performance, which comes from two aspects: 1) device-heterogeneity due to the diverse computational capacity among devices and 2) data-heterogeneity due to the nonidentically distributed data across the network. Prior studies addressing the heterogeneous FL issue, for example, FedProx, lack formalization and it remains an open problem. This work first formalizes the system-heterogeneous FL problem and proposes a new algorithm, called federated local gradient approximation (FedLGA), to address this problem by bridging the divergence of local model updates via gradient approximation. To achieve this, FedLGA provides an alternated Hessian estimation method, which only requires extra linear complexity on the aggregator. Theoretically, we show that with a device-heterogeneous ratio ρ , FedLGA achieves convergence rates on non-i.i.d. distributed FL training data for the nonconvex optimization problems with O ([(1+ρ)/√{ENT}] + 1/T) and O ([(1+ρ)√E/√{TK}] + 1/T) for full and partial device participation, respectively, where E is the number of local learning epoch, T is the number of total communication round, N is the total device number, and K is the number of the selected device in one communication round under partially participation scheme. The results of comprehensive experiments on multiple datasets indicate that FedLGA can effectively address the system-heterogeneous problem and outperform current FL methods. Specifically, the performance against the CIFAR-10 dataset shows that, compared with FedAvg, FedLGA improves the model's best testing accuracy from 60.91% to 64.44%.

Learning from a sequence of tasks for a lifetime is essential for an agent toward artificial general intelligence. Despite the explosion of this research field in recent years, most work focuses on the well-known catastrophic forgetting issue. In contrast, this work aims to explore knowledge-transferable lifelong learning without storing historical data and significant additional computational overhead. We demonstrate that existing data-free frameworks, including regularization-based single-network and structure-based multinetwork frameworks, face a fundamental issue of lifelong learning, named anterograde forgetting, i.e., preserving and transferring memory may inhibit the learning of new knowledge. We attribute it to the fact that the learning network capacity decreases while memorizing historical knowledge and conceptual confusion between the irrelevant old knowledge and the current task. Inspired by the complementary learning theory in neuroscience, we endow artificial neural networks with the ability to continuously learn without forgetting while recalling historical knowledge to facilitate learning new knowledge. Specifically, this work proposes a general framework named cycle memory networks (CMNs). The CMN consists of two individual memory networks to store short- and long-term memories separately to avoid capacity shrinkage and a transfer cell between them. It enables knowledge transfer from the long-term to the short-term memory network to mitigate conceptual confusion. In addition, the memory consolidation mechanism integrates short-term knowledge into the long-term memory network for knowledge accumulation. We demonstrate that the CMN can effectively address the anterograde forgetting on several task-related, task-conflict, class-incremental, and cross-domain benchmarks. Furthermore, we provide extensive ablation studies to verify each framework component. The source codes are available at: https://github.com/GeoX-Lab/CMN.

Federated learning (FL) provides a high efficient decentralized machine learning framework, where the training data remains distributed at remote clients in a network. Though FL enables a privacy-preserving mobile edge computing framework using IoT devices, recent studies have shown that this approach is susceptible to poisoning attacks from the side of remote clients. To address the poisoning attacks on FL, we provide atwo-phasedefense algorithm called${\underline{Lo}cal\ \underline{Ma}licious\ Facto\underline{r}}$(LoMar). In phase I, LoMar scores model updates from each remote client by measuring the relative distribution over their neighbors using a kernel density estimation method. In phase II, an optimal threshold is approximated to distinguish malicious and clean updates from a statistical perspective. Comprehensive experiments on four real-world datasets have been conducted, and the experimental results show that our defense strategy can effectively protect the FL system. Specifically, the defense performance on Amazon dataset under a label-flipping attack indicates that, compared with FG+Krum, LoMar increases the target label testing accuracy from$96.0\%$to$98.8\%$, and the overall averaged testing accuracy from$90.1\%$to$97.0\%$.

Automatic overtaking is a challenging task for self-driving vehicles. Traditional rule-based methods for overtaking in autonomous driving heavily rely on many predefined rules and are difficult to apply in complex driving scenarios. Learning-based methods usually use convolutional networks, recurrent networks, and multilayer perceptrons, etc., to extract features from environments, but they fail to effectively represent geometric and interactive information among traffic participants. Classic graph convolutional networks (GCNs) have the ability of represent graph-structural information but are limited to stable relationship representation due to the fixed adjacency matrix when applied in autonomous driving. In this paper, we propose a novel dynamic graph learning method based on a graph convolutional network with a trainable adjacency matrix (TAM-GCN) to enable the learning of dynamic relationships among different nodes in an ever-changing driving scene. In addition, we develop a planning method for overtaking strategy in autonomous driving, where the proposed TAM-GCN is used to extract the spatial graph-structural features, select appropriate overtaking time, and generate efficient overtaking actions. The proposed model is trained using the imitation learning method. We conduct comprehensive experiments in both closed-loop and open-loop testing in the CARLA simulator and compare our method with state-of-the-art methods. Experimental results demonstrate the proposed method achieves better accuracy, safety and overtaking performance than existing methods.

Multi-server Federated learning (FL) has been considered as a promising solution to address the limited communication resource problem of single-server FL. We consider a typical multi-server FL architecture, where the coverage areas of regional servers may overlap. The key point of this architecture is that the clients located in the overlapping areas update their local models based on the average model of all accessible regional models, which enables indirect model sharing among different regional servers. Due to the complicated network topology, the convergence analysis is much more challenging than in single-server FL. In this paper, we firstly propose a novel MS-FedAvg algorithm for this multi-server FL architecture and analyze its convergence on non-iid datasets for general non-convex settings. Since the number of clients located in each regional server is much less than single-server FL, the bandwidth of each client should be large enough to successfully communicate training models with the server, which indicates that full client participation can work in multi-server FL. Also, we provide the convergence analysis of the partial client participation scheme and develop a new biased partial participation strategy to further accelerate convergence. Our results indicate that the convergence results highly depend on the ratio of the number of clients in each area type to the total number of clients in all three strategies. The extensive experiments show remarkable performance and support our theoretical results.

Federated Learning (FL) is a promising framework for performing privacy-preserving, distributed learning with a set of clients. However, the data distribution among clients often exhibits non-IID, i.e., distribution shift, which makes efficient optimization difficult. To tackle this problem, many FL algorithms focus on mitigating the effects of data heterogeneity across clients by increasing the performance of the global model. However, almost all algorithms leverage Empirical Risk Minimization (ERM) to be the local optimizer, which is easy to make the global model fall into a sharp valley and increase a large deviation of parts of local clients. Therefore, in this paper, we revisit the solutions to the distribution shift problem in FL with a focus on local learning generality. To this end, we propose a general, effective algorithm, \texttt{FedSAM}, based on Sharpness Aware Minimization (SAM) local optimizer, and develop a momentum FL algorithm to bridge local and global models, \texttt{MoFedSAM}. Theoretically, we show the convergence analysis of these two algorithms and demonstrate the generalization bound of \texttt{FedSAM}. Empirically, our proposed algorithms substantially outperform existing FL studies and significantly decrease the learning deviation.

Federated learning (FL) is a new machine learning framework that trains a joint model across a large number of decentralized computing devices. Existing methods, e.g., Federated Averaging (FedAvg), are able to provide an optimization guarantee by synchronously training the joint model, but usually suffer from stragglers, i.e., IoT devices with low computing power or communication bandwidth, especially on optimization problems with non-i.i.d distributed training data. To mitigate the influence of stragglers, this paper presents a novel FL framework, namely Hybrid Federated Learning (HFL), which consists of two major components: synchronous kernel and asynchronous updater that correspondingly addresses the normal devices and stragglers. Unlike traditional synchronous FL frameworks, HFL introduces the asynchronous updater which actively pulls unsyn-chronized and delayed local updates from stragglers to improve the efficiency and effectiveness of FL. Particularly, we propose a novel algorithm, delayed federated-SGD (DF-SGD), to merge the delayed local updates into the joint model. The results of comprehensive experiments under the HFL framework show that DF-SGD outperforms current FL methods.

Soil moisture plays important roles in the land-atmosphere interaction, rainfall–runoff process, climate and weather forecast, and crop yield. An accurate estimation of soil moisture creates opportunities to improve the understanding of soil moisture-related processes in atmospheric sciences, hydrology, and agriculture. Traditional methods to measure soil moisture mainly rely on in-situ measurements which are usually costly and reliant on the use of field instrumentation and point data. A robust remote sensing technique is highly needed to frequently monitor the soil moisture content over a large region or landscape. This paper presents a metric learning-based regression algorithm to learn the mapping function from the input hyperspectral data to the output soil moisture. In particular, the proposed soil moisture estimation method employs the Gaussian process to model the joint distribution of spectral bands over functions. It further incorporates the learning of distance function to find feature subspace for improving soil estimation performance. Experimental results on a real dataset show that the proposed metric learning-based soil moisture estimation method can significantly improve the performance compared to the state-of-the-art soil moisture estimation methods using principle component analysis for feature reduction.

Motion planning is one of the most significant part in autonomous driving. Learning-based motion planning methods attract many researchers’ attention due to the abilities of learning from the environment and directly making decisions from the perception. The deep Q-network, as a popular reinforcement learning method, has achieved great progress in autonomous driving, but these methods seldom use the global path information to handle the issue of directional planning such as making a turning at an intersection since the agent usually learns driving strategies only by the designed reward function, which is difficult to adapt to the driving scenarios of urban roads. Moreover, different motion commands such as the steering wheel and accelerator are associated with each other from classic Q-networks, which easily leads to an unstable prediction of the motion commands since they are independently controlled in a practical driving system. In this paper, a conditional deep Q-network for directional planning is proposed and applied in end-to-end autonomous driving, where the global path is used to guide the vehicle to drive from the origination to the destination. To handle the dependency of different motion commands in Q-networks, we take use of the idea of fuzzy control and develop a defuzzification method to improve the stability of predicting the values of different motion commands. We conduct comprehensive experiments in the CARLA simulator and compare our method with the state-of-the-art methods. Experimental results demonstrate the proposed method achieves better learning performance and driving stability performance than other methods.

Enabling a neural network to sequentially learn multiple tasks is of great significance for expanding the applicability of neural networks in real-world applications. However, artificial neural networks face the well-known problem of catastrophic forgetting. What is worse, the degradation of previously learned skills becomes more severe as the task sequence increases, known as the long-term catastrophic forgetting. It is due to two facts: first, as the model learns more tasks, the intersection of the low-error parameter subspace satisfying for these tasks becomes smaller or even does not exist; second, when the model learns a new task, the cumulative error keeps increasing as the model tries to protect the parameter configuration of previous tasks from interference. Inspired by the memory consolidation mechanism in mammalian brains with synaptic plasticity, we propose a confrontation mechanism in which Adversarial Neural Pruning and synaptic Consolidation (ANPyC) is used to overcome the long-term catastrophic forgetting issue. The neural pruning acts as long-term depression to prune task-irrelevant parameters, while the novel synaptic consolidation acts as long-term potentiation to strengthen task-relevant parameters. During the training, this confrontation achieves a balance in that only crucial parameters remain, and non-significant parameters are freed to learn subsequent tasks. ANPyC avoids forgetting important information and makes the model efficient to learn a large number of tasks. Specifically, the neural pruning iteratively relaxes the current task's parameter conditions to expand the common parameter subspace of the task; the synaptic consolidation strategy, which consists of a structure-aware parameter-importance measurement and an element-wise parameter updating strategy, decreases the cumulative error when learning new tasks. Our approach encourages the synapse to be sparse and polarized, which enables long-term learning and memory. ANPyC exhibits effectiveness and generalization on both image classification and generation tasks with multiple layer perceptron, convolutional neural networks, and generative adversarial networks, and variational autoencoder. The full source code is available at https://github.com/GeoX-Lab/ANPyC.

Obstacle detection and avoidance plays a crucial role in autonomous navigation of unmanned ground vehicles. This becomes more challenging in off-road environments due to the higher probability of finding negative obstacles (e.g., holes, ditches, trenches, etc.) compared with on-road environments. One approach to solve this problem is to avoid the candidate path with a negative obstacle, but in off-road avoiding negative obstacles all the time is not possible. In such cases, the path planner may need to choose a candidate path with a negative obstacle that causes the least amount of damage to the vehicle. To deal better with these types of scenarios, this study introduces a novel approach to perform shape estimation of negative obstacles using LiDAR 3D point cloud data. The dimensions (width, diameter, and depth) and the location (center) of negative obstacles are calculated based on estimated shape. This approach is tested on different terrain types using the Mississippi Autonomous Vehicle Simulation (MAVS).

Anomaly detection from remote sensing images is to detect pixels whose spectral signatures are different from their background. Anomalies are often man-made targets. With such target signatures being unknown, anomaly detection has many important applications, such as water quality monitoring, crop stress surveying, and law enforcement-related uses, where prior information of targets is often unavailable. The key to success is accurate background modeling. Anomaly detection from remote sensing images is challenging because spatial coverage is very large and the background is highly heterogeneous. For pixel-based anomaly detection, computing cost in background modeling and a spatial-convolution-type detection process is very expensive. Thus, parallel and distributed computing is critical in reducing execution time, which can fit the need for real-time or near real-time detection from airborne and spaceborne platforms in support of immediate decision-making. This article reviews the recent advances in anomaly detection from hyperspectral remote sensing images and their implementation using parallel and distributed systems. The classical methods, i.e., the Reed-Xiaoli (RX) algorithm and its variants, including its real-time processing version, are illustrated in commodity graphic processing units (GPUs), cloud, and field-programmable gate array (FPGA) implementations. Practical issues and future development trends are also discussed.

In autonomous driving, many learning-based methods for motion planing have been proposed in literature, which can predict motion commands directly from the sensory data of the environment, but these methods can neither predict multiple motion commands, such as steering angle, accelerator and brake, nor balance errors among different motion commands. In this paper, we propose a deep cascaded neural network for predicting multiple motion commands which can be trained in an end-to-end manner for autonomous driving. The proposed deep cascaded neural network consists of a convolutional neural network (CNN) and three long short-term memory (LSTM) units, fed with images from a front-facing camera installed at the vehicle. As the outputs, the proposed model can predict thee motion planning commands simultaneously including steering angle, acceleration, and brake to enable the autonomous driving. In order to balance errors among different motion commands and improve prediction accuracy, we propose a new network training algorithm, where three independent loss functions are designed to separately update the weights in the three LSTMs connected to three motion commands. We conduct comprehensive experiments using the data from a driving simulator and compare our method with the state-of-the-art methods. Simulation results demonstrate the proposed motion planning model achieves better accuracy performance than other models.

Probabilistic power flow (PPF) calculation is an important power system analysis tool considering the increasing uncertainties. However, existing calculation methods cannot simultaneously achieve high precision and fast calculation, which limits the practical application of the PPF. This article designs a specific architecture of the extreme learning machine (ELM) in a model-driven pattern to extract the power flow features and therefore accelerate the calculation of PPF. ELM is selected because of the unique characteristics of fast training and less intervention. The key challenge is that the learning capability of the ELM for extracting complex features is limited compared with deep neural networks. In this article, we use the physical properties of the power flow model to assist the learning process. To reduce the learning complexity of the power flow features, the feature decomposition and nonlinearity reduction method is proposed to extract the features of the power flow model. An enhanced ELM network architecture is designed. An optimization model for the hidden node parameters is established to improve the learning performance. Based on the proposed model-driven ELM architecture, a fast and accurate PPF calculation method is proposed. The simulations on the IEEE 57-bus and Polish 2383-bus systems demonstrate the effectiveness of the proposed method.

External memory-based neural networks, such as differentiable neural computers (DNCs), have recently gained importance and popularity to solve complex sequential learning tasks that pose challenges to conventional neural networks. However, a trained DNC usually has a low-memory utilization efficiency. This article introduces a variation of DNC architecture with a convertible short-term and long-term memory, named CSLM-DNC. Unlike the memory architecture of the original DNC, the new scheme of short-term and long-term memories offers different importance of memory locations for read and write, and they can be converted over time. This is mainly motivated by the human brain where short-term memory stores large amounts of noisy and unimportant information and decays rapidly, while long-term memory stores important information and lasts for a long time. The conversion of these two types of memory is allowed and is able to be learned according to their reading and writing frequency. We quantitatively and qualitatively evaluate the proposed CSLM-DNC architecture on the tasks of question answering, copy and repeat copy, showing that it can significantly improve memory efficiency and learning performance.

Abrupt change detection is critical to monitor the occurrence of abnormal events from sensor data for situational awareness of complex systems. However, various disturbances and noises applied to the data observations may pose significant challenges to the robustness of many abrupt change detection methods. Recent researches have shown that bilateral filter can acquire outstanding performance on removing noises from images while preserving edge information. In this letter, we propose two improved edge-preserving memory-based cumulative sum (MB-CUSUM) methods that are able to make the abrupt change detection method more robust against noises. Our experimental studies show that the proposed methods can achieve superior performance over state-of-the-art methods to detect abrupt changes, which demonstrates the effectiveness and feasibility of their practical use.

Herein, a generalization of the ordered weighted average (OWA) is put forth relative to pattern recognition. The resultant linear order statistic neuron (LOSN) is unique in that it bridges fuzzy sets, specifically fuzzy data/information aggregation, with neural networks. This article discusses the gradient descent-based optimization and geometric interpretation of the LOSN. An advantage is that the LOSN is an efficient shared weight encoding of N! perceptrons, relative to N inputs. Open source codes are provided to facilitate reproducible research. Experiments are conducted to both validate the method and show its non-linear geometric expression.

Machine learning algorithms have been extensively used to generate complex features for classification task in the hyperspectral images. However, for challenging cases like domain adaptation (DA), these algorithms tend to perform less efficiently. Recently, with the advent of deep learning algorithms, more complex but useful features can be generated for hyperspectral image classification task. However, attention-based feature generation is not explored till now, which has been found to be effective for distinguishing different classes of images than without transferring the parameters. In this paper, we have opted to use attention-based DA based on transferring different levels of attention from a supervisor network to the student network to provide useful but more complex features for improving the overall classification of the DA problem. It has been shown that the proposed attention-based transfer method outperforms the state-of-the-art domain adaptation methods.

This paper proposes a multi-scale rotation-invariant haar-like (MSRI-HL) feature integrated convolutional neural network (MSRIHL-CNN)-based ship detection algorithm of the multiple-target environment in synthetic aperture radar (SAR) imagery. Usually, ship detection includes preprocessing, prescreening, discrimination, and classification. Among them, prescreening and discrimination are the most two important stages so that they catch great intention. Based on our previous work, we propose a truncated-clutter-statistics-based joint, constant false alarm rate (CFAR) detector (TCS-JCFAR) for ship target prescreening in the multiple-target environment. TCS-JCFAR greatly enhances the prescreening rate in the multiple-target environment while achieving a low observed FAR. In the discrimination stage, conventional CNN extracts the deep features (high-level features); however, it will lose the local texture and edge information (low-level features) which are of great significance for target discrimination. Hence, the MSRI-HL features are used to represent the multi-scale, rotation-invariant texture, and edge information that conventional CNN fails to capture. The extracted low-level MSRI-HL features and the high-level deep features are optimally fused to a multi-layered feature vector. Finally, the multi-layered feature vector is fed into a typical support vector machine (SVM) classifier for ship target discrimination. The proposed MSRIHL-CNN combines the low-level texture and edge features and the high-level deep features; moreover, they are optimally fused to fully represent the ship targets. Undoubtedly, MSRIHL-CNN has better discrimination performance. The superiority of the proposed TCS-JCFAR-based prescreener and MSRIHL-CNN-based discriminator is validated on the Chinese Gaofen-3 SAR imagery.

Text categorization proves to be very effective for many language processing applications. Over the years, with the advent of different machine learning entities, the prediction of text categories becomes easier even for a very large document after subtle extraction of features from the documents. Particularly, the learning of sequential features from text data can achieve the state-of-the-art prediction performance. Traditional sequence based deep learning algorithms, namely recurrent neural network (RNN) and its variants, learn and embed features effectively to categorize documents in terms of sequences of a document. However, the memory capacity of these algorithms to track the sequences is limited and less accurate, particularly for the document with a large context. This paper studies an external memory based sequential learning framework to overcome the memory capacity problem. In this work, we propose to employ a multi-layer embedding neural architecture with external memory for the large-scale text categorization problem. It employs long-short-term memory (LSTM) layers as the controller to capture short-term sequential patterns longer and utilizes an external memory to memorize long-term contextual information. An embedded feature layer is introduced to generate one-hot feature vectors from the text. The whole neural architecture is fully differentiable and thus it can be trained in an end-to-end manner. Our experimental results show that the proposed text categorization method outperforms the state-of-the-art methods over a number of large-scale text categorization benchmarks.

Sparse autoencoder is an unsupervised feature extractor and has been widely used in the machine learning and data mining community. However, a sparse hyperparameter has to be determined to balance the trade-off between the reconstruction error and the sparsity of sparse autoencoder. Traditional sparse hyperparameter determination method is time-consuming, especially when the dataset is large. In this paper, we derive a generative model for sparse autoencoder. Based on this model, we derive a formulation to determine the sparse hyperparameter effectively and efficiently. The relationship between the sparse hyperparameter and the average activation of sparse autoencoder hidden units is also presented in this paper. Experimental results and comparative studies over numerous datasets demonstrate the effectiveness of our method to determine the sparse hyperparameter.

This paper develops a new dimensionality reduction method, named Biomimetic Uncorrelated Locality Discriminant Projection (BULDP), for face recognition. It is based on unsupervised discriminant projection and two human bionic characteristics: principle of homology continuity and principle of heterogeneous similarity. With these two human bionic characteristics, we propose a novel adjacency coefficient representation, which does not only capture the category information between different samples, but also reflects the continuity between similar samples and the similarity between different samples. By applying this new adjacency coefficient into the unsupervised discriminant projection, it can be shown that we can transform the original data space into an uncorrelated discriminant subspace. A detailed solution of the proposed BULDP is given based on singular value decomposition. Moreover, we also develop a nonlinear version of our BULDP using kernel functions for nonlinear dimensionality reduction. The performance of the proposed algorithms is evaluated and compared with the state-of-the-art methods on four public benchmarks for face recognition. Experimental results show that the proposed BULDP method and its nonlinear version achieve much competitive recognition performance.

Using dictionary atoms to reconstruct input vectors is of great interest in spare representation. However, a key challenge is how to find a proper dictionary. In this paper, we introduce an active dictionary learning (ADL) method which incorporates active learning criteria to select atoms for dictionary construction with the consideration of both classification and reconstruction errors. Specifically, we apply a sparse representation based classification (SRC) method to calculate the learned dictionary and use the classification accuracy and the reconstruction error to evaluate the proposed dictionary learning method. In our experiments, we compare the performance of our proposed dictionary learning method with many other methods, including unsupervised dictionary learning and whole-training-data dictionary, on several UCI data sets and the Extended Yale B face data set. The superior performance demonstrates the effectiveness of the proposed method.

Data intensive analysis is the major challenge in smart cities because of the ubiquitous deployment of various kinds of sensors. The natural characteristic of geodistribution requires a new computing paradigm to offer location-awareness and latency-sensitive monitoring and intelligent control. Fog Computing that extends the computing to the edge of network, fits this need. In this paper, we introduce a hierarchical distributed Fog Computing architecture to support the integration of massive number of infrastructure components and services in future smart cities. To secure future communities, it is necessary to integrate intelligence in our Fog Computing architecture, e.g., to perform data representation and feature extraction, to identify anomalous and hazardous events, and to offer optimal responses and controls. We analyze case studies using a smart pipeline monitoring system based on fiber optic sensors and sequential learning algorithms to detect events threatening pipeline safety. A working prototype was constructed to experimentally evaluate event detection performance of the recognition of 12 distinct events. These experimental results demonstrate the feasibility of the system's city-wide implementation in the future.

Feature selection aims to gain relevant features for improved classification performance and remove redundant features for reduced computational cost. How to balance these two factors is a problem especially when the categorical labels are costly to obtain. In this paper, we address this problem using semisupervised learning method and propose a max-relevance and min-redundancy criterion based on Pearson's correlation (RRPC) coefficient. This new method uses the incremental search technique to select optimal feature subsets. The new selected features have strong relevance to the labels in supervised manner, and avoid redundancy to the selected feature subsets under unsupervised constraints. Comparative studies are performed on binary data and multicategory data from benchmark data sets. The results show that the RRPC can achieve a good balance between relevance and redundancy in semisupervised feature selection. We also compare the RRPC with classic supervised feature selection criteria (such as mRMR and Fisher score), unsupervised feature selection criteria (such as Laplacian score), and semisupervised feature selection criteria (such as sSelect and locality sensitive). Experimental results demonstrate the effectiveness of our method.

Understanding human mobility is important for the development of intelligent mobile service robots as it can provide prior knowledge and predictions of human distribution for robot-assisted activities. In this paper, we propose a probabilistic method to model human motion behaviors which is determined by both internal and external factors in an indoor environment. While the internal factors are represented by the individual preferences, aims and interests, the external factors are indicated by the stimulation of the environment. We model the randomness of human macro-level movement, e.g., the probability of visiting a specific place and staying time, under the Bayesian framework, considering the influence of both internal and external variables. We use two case studies in a shopping mall and in a college student dorm building to show the effectiveness of our proposed probabilistic human mobility model. Real surveillance camera data are used to validate the proposed model together with survey data in the case study of student dorm.

The threat of false data injection (FDI) attacks have raised wide interest in the research and development of smart grid security. This paper presents a comparative study on the utilization of supervised learning classifiers to detect direct and stealth FDI attacks in the smart grid. A detailed formulation of the problem for detection with classifiers is first described with proper assumptions and justifications. Three widely used supervised learning (SL) based classifiers are chosen to design corresponding FDI detectors. The performance are tested against false measurement data (direct FDI attack) and false state data (stealth FDI attack) on both balanced and imbalanced cases, with consideration of the influence of FDI resources and magnitudes. Simulations on IEEE 30-bus system have shown that the SL based detectors can effectively detect both direct and stealth FDI attacks, especially for the more severe attacks with large amount or magnitude of compromised measurements.

In this paper, we present a novel exponentially embedded families (EEF) based classification method, in which the probability density function (PDF) on raw data is estimated from the PDF on features. With the PDF construction, we show that class-specific features can be used in the proposed classification method, instead of a common feature subset for all classes as used in conventional approaches. We apply the proposed EEF classifier for text categorization as a case study and derive an optimal Bayesian classification rule with class-specific feature selection based on the Information Gain score. The promising performance on real-life data sets demonstrates the effectiveness of the proposed approach and indicates its wide potential applications.

A large number of injuries or deaths may occur when an emergency happens in a crowded public place. The congestion at exits may slow down the egress rate due to the effect of “faster-is-slower”. This inspires us to study how human behavior dynamically changes over time at an emergency in a complex indoor environment. In this paper, we refer the panic of evacuees to their perception of the threat and propose a panic propagation model to model how crowd panic changes during evacuation at an emergency. Combined with the existing social force model, our panic model interprets the self-driven force and interactive forces with others in human mobility. To improve evacuation efficiency, robots are introduced to guide evacuees to escape. Using dynamic environment information, we design an evacuation exit selection algorithm where the optimal exit is automatically selected by the robot with the minimum escape time. In our experiments, a real shopping mall is examined, and the dynamic behavior of panicked evacuees is simulated with the proposed panic model. The evacuation performance of using emergency evacuation robots is evaluated. The improvement of evacuation efficiency validates the effectiveness of our robot-assisted evacuation system.

In this paper, we present a Bayesian classification approach for automatic text categorization using class-specific features. Unlike conventional text categorization approaches, our proposed method selects a specific feature subset for each class. To apply these class-specific features for classification, we follow Baggenstoss's PDF Projection Theorem (PPT) to reconstruct the PDFs in raw data space from the class-specific PDFs in low-dimensional feature subspace, and build a Bayesian classification rule. One noticeable significance of our approach is that most feature selection criteria, such as Information Gain (IG) and Maximum Discrimination (MD), can be easily incorporated into our approach. We evaluate our method's classification performance on several real-world benchmarks, compared with the state-of-the-art feature selection approaches. The superior results demonstrate the effectiveness of the proposed approach and further indicate its wide potential applications in data mining.

Automated feature selection is important for text categorization to reduce feature size and to speed up learning process of classifiers. In this paper, we present a novel and efficient feature selection framework based on the Information Theory, which aims to rank the features with their discriminative capacity for classification. We first revisit two information measures: Kullback-Leibler divergence and Jeffreys divergence for binary hypothesis testing, and analyze their asymptotic properties relating to type I and type II errors of a Bayesian classifier. We then introduce a new divergence measure, called Jeffreys-Multi-Hypothesis (JMH) divergence, to measure multi-distribution divergence for multi-class classification. Based on the JMH-divergence, we develop two efficient feature selection methods, termed maximum discrimination ($MD$) and methods, for text categorization. The promising results of extensive experiments demonstrate the effectiveness of the proposed approaches.

In imbalanced learning, most standard classification algorithms usually fail to properly represent data distribution and provide unfavorable classification performance. More specifically, the decision rule of minority class is usually weaker than majority class, leading to many misclassification of expensive minority class data. Motivated by our previous work ADASYN [1], this paper presents a novel kernel based adaptive synthetic over-sampling approach, named KernelADASYN, for imbalanced data classification problems. The idea is to construct an adaptive over-sampling distribution to generate synthetic minority class data. The adaptive over-sampling distribution is first estimated with kernel density estimation methods and is further weighted by the difficulty level for different minority class data. The classification performance of our proposed adaptive over-sampling approach is evaluated on several real-life benchmarks, specifically on medical and healthcare applications. The experimental results show the competitive classification performance for many real-life imbalanced data classification problems.

We introduce a new parallel computing and communication architecture, Reflex-Tree, with massive sensing, data processing, and control functions suitable for future smart cities. The central feature of the proposed Reflex-Tree architecture is inspired by a fundamental element of the human nervous system: reflex arcs, the neuromuscular reactions and instinctive motions of a part of the body in response to urgent situations. At the bottom level of the Reflex-Tree (layer 4), novel sensing devices are proposed that are controlled by low power processing elements. These "leaf" nodes are then connected to new classification engines based on machine learning techniques, including support vector machines (SVM), to form the third layer. The next layer up consists of servers that provide accurate control decisions via multi-layer adaptive learning and spatial-temporal association, before they are connected to the top level cloud where complex system behavior analysis is performed. Our multi-layered architecture mimics human neural circuits to achieve the high levels of parallelization and scalability required for efficient city-wide monitoring and feedback. To demonstrate the utility of our architecture, we present the design, implementation, and experimental evaluation of a prototype Reflex-Tree. City power supply network and gas pipeline management scenarios are used to drive our prototype as case studies. We show the effectiveness for several levels of the architecture and discuss the feasibility of implementation.

In this paper, we extend the exponentially embedded family (EEF), a new approach to model order estimation and probability density function construction originally proposed by Kay in 2005, to multivariate pattern recognition. Specifically, a parametric classifier rule based on the EEF is developed, in which we construct a distribution for each class based on a reference distribution. The proposed method can address different types of classification problems in either a data-driven manner or a model-driven manner. In this paper, we demonstrate its effectiveness with examples of synthetic data classification and real-life data classification in a data-driven manner and the example of power quality disturbance classification in a model-driven manner. To evaluate the classification performance of our approach, the Monte-Carlo method is used in our experiments. The promising experimental results indicate many potential applications of the proposed method.

Previous works on outdoor traffic sign recognition and classification have been demonstrated useful to the driver assistant system and the possibility to the autonomous vehicles. This motivates our research on the assistance for visual impairment or visual disabled pedestrians in the indoor environment. In this paper, we build an indoor sign database and investigate the recognition and classification for the indoor sign problem. We adopt the classical techniques on extracting the features, including the principle component analysis (PCA), dense scale invariant feature transform (DSIFT), histogram of oriented gradients (HOG), and conduct the state-of-art classification techniques, such as the neural network (NN), support vector machine (SVM) and k-nearest neighbors (KNN). We provide the experimental results on this newly built database and also discuss the insight for the possibility of indoor navigation for the blind or visual-disabled people.

The parametric classifiers trained with the Bayesian rule are usually more accurate than the non-parametric classifiers such as nearest neighbors, neural network and support vector machine, when the class-conditional densities of distribution models are known except for some of their parameters and the training data is abundant. However, the parametric classifiers would perform poorly if these class-conditional densities are unknown and the assumed distribution models are inaccurate. In this paper, we propose a hybrid classification method for the data with partially known distribution models where only the distribution models of some classes are known. For this partial models case, the proposed hybrid classifier makes the best use of knowledge of known distribution models with Bayesian interference, while both purely parametric and non-parametric classifiers would lose a specific predictive capacity for classification. Theoretical proofs and experimental results show that the proposed hybrid classifier has much better performance than these purely parametric and non-parametric classifiers for the data with partial models.