In today's connected world,the generation of massive streaming data across diverse domains has become commonplace.In the presence of concept drift,class imbalance,label scarcity,and new class emergence,these chall...In today's connected world,the generation of massive streaming data across diverse domains has become commonplace.In the presence of concept drift,class imbalance,label scarcity,and new class emergence,these challenges jointly degrade representation stability,bias learning toward outdated distributions,and reduce the resilience and reliability of detection in dynamic environments.This paper proposes a streaming classincremental learning(SCIL)framework to address these issues.The SCIL framework integrates an autoencoder(AE)with a multi-layer perceptron for multi-class prediction,employs a dual-loss strategy(classification and reconstruction)for prediction and new class detection,uses corrected pseudo-labels for online training,manages classes with queues,and applies oversampling to handle imbalance.The rationale behind the method's structure is elucidated through ablation studies,and a comprehensive experimental evaluation is performed using both real-world and synthetic datasets that feature class imbalance,incremental classes,and concept drifts.Our results demonstrate that SCIL outperforms strong baselines and state-of-the-art methods.In line with our commitment to Open Science,we make our code and datasets available to the community.展开更多
Modern intelligent systems,such as autonomous vehicles and face recognition,must continuously adapt to new scenarios while preserving their ability to handle previously encountered situations.However,when neural netwo...Modern intelligent systems,such as autonomous vehicles and face recognition,must continuously adapt to new scenarios while preserving their ability to handle previously encountered situations.However,when neural networks learn new classes sequentially,they suffer from catastrophic forgetting—the tendency to lose knowledge of earlier classes.This challenge,which lies at the core of class-incremental learning,severely limits the deployment of continual learning systems in real-world applications with streaming data.Existing approaches,including rehearsalbased methods and knowledge distillation techniques,have attempted to address this issue but often struggle to effectively preserve decision boundaries and discriminative features under limited memory constraints.To overcome these limitations,we propose a support vector-guided framework for class-incremental learning.The framework integrates an enhanced feature extractor with a Support Vector Machine classifier,which generates boundary-critical support vectors to guide both replay and distillation.Building on this architecture,we design a joint feature retention strategy that combines boundary proximity with feature diversity,and a Support Vector Distillation Loss that enforces dual alignment in decision and semantic spaces.In addition,triple attention modules are incorporated into the feature extractor to enhance representation power.Extensive experiments on CIFAR-100 and Tiny-ImageNet demonstrate effective improvements.On CIFAR-100 and Tiny-ImageNet with 5 tasks,our method achieves 71.68%and 58.61%average accuracy,outperforming strong baselines by 3.34%and 2.05%.These advantages are consistently observed across different task splits,highlighting the robustness and generalization of the proposed approach.Beyond benchmark evaluations,the framework also shows potential in few-shot and resource-constrained applications such as edge computing and mobile robotics.展开更多
Continual learning aims to empower a model to learn new tasks continuously while reducing forgetting to retain previously learnt knowledge.In the context of receiving streaming data that are not constrained by the ind...Continual learning aims to empower a model to learn new tasks continuously while reducing forgetting to retain previously learnt knowledge.In the context of receiving streaming data that are not constrained by the independent and identically distributed(IID)assumption,continual learning efficiently transforms and leverages previously learnt knowledge through various methodologies and completes the learning of new tasks.The generalisation performance and learning efficiency of the model are enhanced in a sequence of tasks.However,the class imbalance in continual learning scenarios critically undermines model performance.In particular,in the class-incremental scenario,the class imbalance results in a bias towards new task classes while degrading the performance on previous learnt classes,leading to catastrophic forgetting.In this paper,a novel method based on balanced contrast is proposed to solve the class-incremental continual learning.The method utilises gradient balancing to mitigate the impact of class imbalance in the class-incremental scenario.The method leverages contrastive learning and gradient modifications to facilitate balanced processing of data across different classes in continual learning.The method proposed in this paper surpasses the existing baseline approaches in the class-incremental learning scenario on standard image datasets such as CIFAR-100,CIFAR-10 and mini-ImageNet.The research results reveal that the proposed method effectively mitigates catastrophic forgetting of previously learnt classes,markedly improving the efficacy of continual learning and offering a powerful solution for further advancing continual learning performance.展开更多
Deep Convolution Neural Networks(DCNNs)can capture discriminative features from large datasets.However,how to incrementally learn new samples without forgetting old ones and recognize novel classes that arise in the d...Deep Convolution Neural Networks(DCNNs)can capture discriminative features from large datasets.However,how to incrementally learn new samples without forgetting old ones and recognize novel classes that arise in the dynamically changing world,e.g.,classifying newly discovered fish species,remains an open problem.We address an even more challenging and realistic setting of this problem where new class samples are insufficient,i.e.,Few-Shot Class-Incremental Learning(FSCIL).Current FSCIL methods augment the training data to alleviate the overfitting of novel classes.By contrast,we propose Filter Bank Networks(FBNs)that augment the learnable filters to capture fine-detailed features for adapting to future new classes.In the forward pass,FBNs augment each convolutional filter to a virtual filter bank containing the canonical one,i.e.,itself,and multiple transformed versions.During back-propagation,FBNs explicitly stimulate fine-detailed features to emerge and collectively align all gradients of each filter bank to learn the canonical one.FBNs capture pattern variants that do not yet exist in the pretraining session,thus making it easy to incorporate new classes in the incremental learning phase.Moreover,FBNs introduce model-level prior knowledge to efficiently utilize the limited few-shot data.Extensive experiments on MNIST,CIFAR100,CUB200,andMini-ImageNet datasets show that FBNs consistently outperformthe baseline by a significantmargin,reporting new state-of-the-art FSCIL results.In addition,we contribute a challenging FSCIL benchmark,Fishshot1K,which contains 8261 underwater images covering 1000 ocean fish species.The code is included in the supplementary materials.展开更多
Federated learning(FL)enables collaborative model training among participants while guaranteeing the privacy of raw data.Mainstream FL methodologies overlook the dynamic nature of real-world data,particularly its tend...Federated learning(FL)enables collaborative model training among participants while guaranteeing the privacy of raw data.Mainstream FL methodologies overlook the dynamic nature of real-world data,particularly its tendency to grow in volume and diversify in classes over time.This oversight results in FL methods suffering from catastrophic forgetting,where the trained models inadvertently discard previously learned information upon assimilating new data.In response to this challenge,we propose a novel federated class-incremental learning(FCIL)method,named Federated Classincremental Learning with New-Class Augmented Self-Distillation(FedCLASS).The core of FedCLASS is to enrich the class scores of historical models with new class scores predicted by current models and utilize the combined knowledge for self-distillation,enabling a more sufficient and precise knowledge transfer from historical models to current models.Theoretical analyses demonstrate that FedCLASS stands on reliable foundations,considering the scores of old classes predicted by historical models as conditional probabilities in the absence of new classes,and the scores of new classes predicted by current models as the conditional probabilities of class scores derived from historical models.Empirical experiments demonstrate the superiority of FedCLASS over four baseline algorithms in reducing average forgetting rate and boosting global accuracy.展开更多
In wireless sensor networks,ensuring communication security via specific emitter identification(SEI)is crucial.However,existing SEI methods are limited to closed-set scenarios and lack the ability to detect unknown de...In wireless sensor networks,ensuring communication security via specific emitter identification(SEI)is crucial.However,existing SEI methods are limited to closed-set scenarios and lack the ability to detect unknown devices and perform classincremental training.This study proposes a class-incremental open-set SEI approach.The open-set SEI model calculates radiofrequency fingerprints(RFFs)prototypes for known signals and employs a self-attention mechanism to enhance their discriminability.Detection thresholds are set through Gaussian fitting for each class.For class-incremental learning,the algorithm freezes the parameters of the previously trained model to initialize the new model.It designs specific losses:the RFFs extraction distribution difference loss and the prototype transformation distribution difference loss,which force the new model to retain old knowledge while learning new knowledge.The training loss enables learning of new class RFFs.Experimental results demonstrate that the open-set SEI model achieves state-of-theart performance and strong noise robustness.Moreover,the class-incremental learning algorithm effectively enables the model to retain old device RFFs knowledge,acquire new device RFFs knowledge,and detect unknown devices simultaneously.展开更多
1 Introduction Current continual learning methods[1–4]can utilize labeled data to alleviate catastrophic forgetting effectively.However,obtaining labeled samples can be difficult and tedious as it may require expert ...1 Introduction Current continual learning methods[1–4]can utilize labeled data to alleviate catastrophic forgetting effectively.However,obtaining labeled samples can be difficult and tedious as it may require expert knowledge.In many practical application scenarios,labeled and unlabeled samples exist simultaneously,with more unlabeled than labeled samples in streaming data[5,6].Unfortunately,existing class-incremental learning methods face limitations in effectively utilizing unlabeled data,thereby impeding their performance in incremental learning scenarios.展开更多
Continual learning(CL)studies the problem of learning to accumulate knowledge over time from a stream of data.A crucial challenge is that neural networks suffer from performance degradation on previously seen data,kno...Continual learning(CL)studies the problem of learning to accumulate knowledge over time from a stream of data.A crucial challenge is that neural networks suffer from performance degradation on previously seen data,known as catastrophic forgetting,due to allowing parameter sharing.In this work,we consider a more practical online class-incremental CL setting,where the model learns new samples in an online manner and may continuously experience new classes.Moreover,prior knowledge is unavailable during training and evaluation.Existing works usually explore sample usages from a single dimension,which ignores a lot of valuable supervisory information.To better tackle the setting,we propose a novel replay-based CL method,which leverages multi-level representations produced by the intermediate process of training samples for replay and strengthens supervision to consolidate previous knowledge.Specifically,besides the previous raw samples,we store the corresponding logits and features in the memory.Furthermore,to imitate the prediction of the past model,we construct extra constraints by leveraging multi-level information stored in the memory.With the same number of samples for replay,our method can use more past knowledge to prevent interference.We conduct extensive evaluations on several popular CL datasets,and experiments show that our method consistently outperforms state-of-the-art methods with various sizes of episodic memory.We further provide a detailed analysis of these results and demonstrate that our method is more viable in practical scenarios.展开更多
The uplift resistance of the soil overlying shield tunnels significantly impacts their anti-floating stability.However,research on uplift resistance concerning special-shaped shield tunnels is limited.This study combi...The uplift resistance of the soil overlying shield tunnels significantly impacts their anti-floating stability.However,research on uplift resistance concerning special-shaped shield tunnels is limited.This study combines numerical simulation with machine learning techniques to explore this issue.It presents a summary of special-shaped tunnel geometries and introduces a shape coefficient.Through the finite element software,Plaxis3D,the study simulates six key parameters—shape coefficient,burial depth ratio,tunnel’s longest horizontal length,internal friction angle,cohesion,and soil submerged bulk density—that impact uplift resistance across different conditions.Employing XGBoost and ANN methods,the feature importance of each parameter was analyzed based on the numerical simulation results.The findings demonstrate that a tunnel shape more closely resembling a circle leads to reduced uplift resistance in the overlying soil,whereas other parameters exhibit the contrary effects.Furthermore,the study reveals a diminishing trend in the feature importance of buried depth ratio,internal friction angle,tunnel longest horizontal length,cohesion,soil submerged bulk density,and shape coefficient in influencing uplift resistance.展开更多
Hand gesture recognition(HGR)plays a vital role in human-computer interaction.The integration of high-density surface electromyography(HD-sEMG)and deep neural networks(DNNs)has significantly improved the robustness an...Hand gesture recognition(HGR)plays a vital role in human-computer interaction.The integration of high-density surface electromyography(HD-sEMG)and deep neural networks(DNNs)has significantly improved the robustness and accuracy of HGR systems.These methods are typically effective for a fixed set of trained gestures.However,the need for new gesture classes over time poses a challenge.Introducing new classes to DNNs can lead to a substantial decrease in accuracy for previously learned tasks,a phenomenon known as“catastrophic forgetting,”especially when the training data for earlier tasks is not retained and retrained.This issue is exacerbated in embedded devices with limited storage,which struggle to store the large-scale data of HD-sEMG.Classincremental learning(CIL)is an effective method to reduce catastrophic forgetting.However,existing CIL methods for HGR rarely focus on reducing memory load.To address this,we propose a memory-friendly CIL method for HGR using HD-sEMG.Our approach includes a lightweight convolutional neural network,named SeparaNet,for feature representation learning,coupled with a nearest-mean-of-exemplars classifier for classifi-cation.We introduce a priority exemplar selection algorithm inspired by the herding effect to maintain a manageable set of exemplars during training.Furthermore,a task-equal-weight exemplar sampling strategy is proposed to effectively reduce memory load while preserving high recognition performance.Experimental results on two datasets demonstrate that our method significantly reduces the number of retained exemplars to only a quarter of that required by other CIL methods,accounting for less than 5%of the total samples,while still achieving comparable average accuracy.展开更多
Countries around the world have been making efforts to reduce pollutant emissions. However, the response of global black carbon(BC) aging to emission changes remains unclear. Using the Community Atmosphere Model versi...Countries around the world have been making efforts to reduce pollutant emissions. However, the response of global black carbon(BC) aging to emission changes remains unclear. Using the Community Atmosphere Model version 6 with a machine-learning-integrated four-mode version of the Modal Aerosol Module, we quantify global BC aging responses to emission reductions for 2011–2018 and for 2050 and 2100 under carbon neutrality. During 2011–18, global trends in BC aging degree(mass ratio of coatings to BC, R_(BC)) exhibited marked regional disparities, with a significant increase in China(5.4% yr^(-1)), which contrasts with minimal changes in the USA, Europe, and India. The divergence is attributed to opposing trends in secondary organic aerosol(SOA) and sulfate coatings, driven by regional changes in the emission ratios of corresponding coating precursors to BC(volatile organic compounds-VOCs/BC and SO_(2)/BC). Projections under carbon neutrality reveal that R_(BC) will increase globally by 47%(118%) in 2050(2100), with strong convergent increases expected across major source regions. The R_(BC) increase, primarily driven by enhanced SOA coatings due to sharper BC reductions relative to VOCs, will enhance the global BC mass absorption cross-section(MAC) by 11%(17%) in 2050(2100).Consequently, although the global BC burden will decline sharply by 60%(76%), the enhanced MAC partially offsets the magnitude of the decline in the BC direct radiative effect, resulting in the moderation of global BC DRE decreases to 88%(92%) of the BC burden reductions in 2050(2100). This study highlights the globally enhanced BC aging and light absorption capacity under carbon neutrality, thereby partly offsetting the impact of BC direct emission reductions on future changes in BC radiative effects globally.展开更多
Underwater pipeline inspection plays a vital role in the proactive maintenance and management of critical marine infrastructure and subaquatic systems.However,the inspection of underwater pipelines presents a challeng...Underwater pipeline inspection plays a vital role in the proactive maintenance and management of critical marine infrastructure and subaquatic systems.However,the inspection of underwater pipelines presents a challenge due to factors such as light scattering,absorption,restricted visibility,and ambient noise.The advancement of deep learning has introduced powerful techniques for processing large amounts of unstructured and imperfect data collected from underwater environments.This study evaluated the efficacy of the You Only Look Once(YOLO)algorithm,a real-time object detection and localization model based on convolutional neural networks,in identifying and classifying various types of pipeline defects in underwater settings.YOLOv8,the latest evolution in the YOLO family,integrates advanced capabilities,such as anchor-free detection,a cross-stage partial network backbone for efficient feature extraction,and a feature pyramid network+path aggregation network neck for robust multi-scale object detection,which make it particularly well-suited for complex underwater environments.Due to the lack of suitable open-access datasets for underwater pipeline defects,a custom dataset was captured using a remotely operated vehicle in a controlled environment.This application has the following assets available for use.Extensive experimentation demonstrated that YOLOv8 X-Large consistently outperformed other models in terms of pipe defect detection and classification and achieved a strong balance between precision and recall in identifying pipeline cracks,rust,corners,defective welds,flanges,tapes,and holes.This research establishes the baseline performance of YOLOv8 for underwater defect detection and showcases its potential to enhance the reliability and efficiency of pipeline inspection tasks in challenging underwater environments.展开更多
Cyclohexene is an important raw material in the production of nylon.Selective hydrogenation of benzene is a key method for preparing cyclohexene.However,the Ru catalysts used in current industrial processes still face...Cyclohexene is an important raw material in the production of nylon.Selective hydrogenation of benzene is a key method for preparing cyclohexene.However,the Ru catalysts used in current industrial processes still face challenges,including high metal usage,high process costs,and low cyclohexene yield.This study utilizes existing literature data combined with machine learning methods to analyze the factors influencing benzene conversion,cyclohexene selectivity,and yield in the benzene hydrogenation to cyclohexene reaction.It constructs predictive models based on XGBoost and Random Forest algorithms.After analysis,it was found that reaction time,Ru content,and space velocity are key factors influencing cyclohexene yield,selectivity,and benzene conversion.Shapley Additive Explanations(SHAP)analysis and feature importance analysis further revealed the contribution of each variable to the reaction outcomes.Additionally,we randomly generated one million variable combinations using the Dirichlet distribution to attempt to predict high-yield catalyst formulations.This paper provides new insights into the application of machine learning in heterogeneous catalysis and offers some reference for further research.展开更多
The electrocatalytic reduction of nitric oxide for ammonia synthesis(NORR)is a key green energy conversion technology.Its efficiency relies on high-performance electrocatalysts to enhance both ammonia yield(Y_(NH3))an...The electrocatalytic reduction of nitric oxide for ammonia synthesis(NORR)is a key green energy conversion technology.Its efficiency relies on high-performance electrocatalysts to enhance both ammonia yield(Y_(NH3))and Faradaic efficiency(F_(NH3)).However,conventional experimental methods for screening high-activity NORR catalysts often entail high resource consumption and time costs.Machine learning combined with SHAP feature analysis was employed to establish a stacked ensemble model that integrates multiple algorithms,to allow for a systematic investigation of the key descriptors governing NORR performance based on an experimental dataset.Evaluation of eight model algorithms revealed that the Stacked-SVR model achieved an R^(2)of 0.9223 and an RMSE of 0.0608 for predicting on the test set,whereas the Stacked-RF model achieved an R^(2)of 0.9042 and an RMSE of 0.0900 for predicting.The stacked ensemble model integrates the strengths of individual algorithms and demonstrates strong NORR prediction performance while avoiding overfitting.SHAP feature analysis results revealed that the Cu content in the catalyst composition has the most significant impact on catalytic performance.Moreover,the combination of the wet chemical reduction synthesis,a carbon fiber(CF)conductive substrate,and HCl electrolyte is more favorable for enhancing catalytic activity.Additionally,moderately lowering the working potential,controlling the electrolyte volume at low to medium levels,reducing catalyst loading,and increasing electrolyte concentration were found to synergistically enhance both and.展开更多
Federated Learning(FL)has become a leading decentralized solution that enables multiple clients to train a model in a collaborative environment without directly sharing raw data,making it suitable for privacy-sensitiv...Federated Learning(FL)has become a leading decentralized solution that enables multiple clients to train a model in a collaborative environment without directly sharing raw data,making it suitable for privacy-sensitive applications such as healthcare,finance,and smart systems.As the field continues to evolve,the research field has become more complex and scattered,covering different system designs,training methods,and privacy techniques.This survey is organized around the three core challenges:how the data is distributed,how models are synchronized,and how to defend against attacks.It provides a structured and up-to-date review of FL research from 2023 to 2025,offering a unified taxonomy that categorizes works by data distribution(Horizontal FL,Vertical FL,Federated Transfer Learning,and Personalized FL),training synchronization(synchronous and asynchronous FL),optimization strategies,and threat models(data leakage and poisoning attacks).In particular,we summarize the latest contributions in Vertical FL frameworks for secure multi-party learning,communication-efficient Horizontal FL,and domain-adaptive Federated Transfer Learning.Furthermore,we examine synchronization techniques addressing system heterogeneity,including straggler mitigation in synchronous FL and staleness management in asynchronous FL.The survey covers security threats in FL,such as gradient inversion,membership inference,and poisoning attacks,as well as their defense strategies that include privacy-preserving aggregation and anomaly detection.The paper concludes by outlining unresolved issues and highlighting challenges in handling personalized models,scalability,and real-world adoption.展开更多
Split Learning(SL)has been promoted as a promising collaborative machine learning technique designed to address data privacy and resource efficiency.Specifically,neural networks are divided into client and server subn...Split Learning(SL)has been promoted as a promising collaborative machine learning technique designed to address data privacy and resource efficiency.Specifically,neural networks are divided into client and server subnetworks in order to mitigate the exposure of sensitive data and reduce the overhead on client devices,thereby making SL particularly suitable for resource-constrained devices.Although SL prevents the direct transmission of raw data,it does not alleviate entirely the risk of privacy breaches.In fact,the data intermediately transmitted to the server sub-model may include patterns or information that could reveal sensitive data.Moreover,achieving a balance between model utility and data privacy has emerged as a challenging problem.In this article,we propose a novel defense approach that combines:(i)Adversarial learning,and(ii)Network channel pruning.In particular,the proposed adversarial learning approach is specifically designed to reduce the risk of private data exposure while maintaining high performance for the utility task.On the other hand,the suggested channel pruning enables the model to adaptively adjust and reactivate pruned channels while conducting adversarial training.The integration of these two techniques reduces the informativeness of the intermediate data transmitted by the client sub-model,thereby enhancing its robustness against attribute inference attacks without adding significant computational overhead,making it wellsuited for IoT devices,mobile platforms,and Internet of Vehicles(IoV)scenarios.The proposed defense approach was evaluated using EfficientNet-B0,a widely adopted compact model,along with three benchmark datasets.The obtained results showcased its superior defense capability against attribute inference attacks compared to existing state-of-the-art methods.This research’s findings demonstrated the effectiveness of the proposed channel pruning-based adversarial training approach in achieving the intended compromise between utility and privacy within SL frameworks.In fact,the classification accuracy attained by the attackers witnessed a drastic decrease of 70%.展开更多
The Internet of Vehicles,or IoV,is expected to lessen pollution,ease traffic,and increase road safety.IoV entities’interconnectedness,however,raises the possibility of cyberattacks,which can have detrimental effects....The Internet of Vehicles,or IoV,is expected to lessen pollution,ease traffic,and increase road safety.IoV entities’interconnectedness,however,raises the possibility of cyberattacks,which can have detrimental effects.IoV systems typically send massive volumes of raw data to central servers,which may raise privacy issues.Additionally,model training on IoV devices with limited resources normally leads to slower training times and reduced service quality.We discuss a privacy-preserving Federated Split Learning with Tiny Machine Learning(TinyML)approach,which operates on IoV edge devices without sharing sensitive raw data.Specifically,we focus on integrating split learning(SL)with federated learning(FL)and TinyML models.FL is a decentralisedmachine learning(ML)technique that enables numerous edge devices to train a standard model while retaining data locally collectively.The article intends to thoroughly discuss the architecture and challenges associated with the increasing prevalence of SL in the IoV domain,coupled with FL and TinyML.The approach starts with the IoV learning framework,which includes edge computing,FL,SL,and TinyML,and then proceeds to discuss how these technologies might be integrated.We elucidate the comprehensive operational principles of Federated and split learning by examining and addressingmany challenges.We subsequently examine the integration of SL with FL and various applications of TinyML.Finally,exploring the potential integration of FL and SL with TinyML in the IoV domain is referred to as FSL-TM.It is a superior method for preserving privacy as it conducts model training on individual devices or edge nodes,thereby obviating the necessity for centralised data aggregation,which presents considerable privacy threats.The insights provided aim to help both researchers and practitioners understand the complicated terrain of FL and SL,hence facilitating advancement in this swiftly progressing domain.展开更多
Knowledge distillation has become a standard technique for compressing large language models into efficient student models,but existing methods often struggle to balance prediction accuracy with explanation quality.Re...Knowledge distillation has become a standard technique for compressing large language models into efficient student models,but existing methods often struggle to balance prediction accuracy with explanation quality.Recent approaches such as Distilling Step-by-Step(DSbS)introduce explanation supervision,yet they apply it in a uniform manner that may not fully exploit the different learning dynamics of prediction and explanation.In this work,we propose a task-structured curriculum learning(TSCL)framework that structures training into three sequential phases:(i)prediction-only,to establish stable feature representations;(ii)joint prediction-explanation,to align task outputs with rationale generation;and(iii)explanation-only,to refine the quality of rationales.This design provides a simple but effective modification to DSbS,requiring no architectural changes and adding negligible training cost.We justify the phase scheduling with ablation studies and convergence analysis,showing that an initial prediction-heavy stage followed by a balanced joint phase improves both stability and explanation alignment.Extensive experiments on five datasets(e-SNLI,ANLI,CommonsenseQA,SVAMP,and MedNLI)demonstrate that TSCL consistently outperforms strong baselines,achieving gains of+1.7-2.6 points in accuracy and 0.8-1.2 in ROUGE-L,corresponding to relative error reductions of up to 21%.Beyond lexical metrics,human evaluation and ERASERstyle faithfulness diagnostics confirm that TSCL produces more faithful and informative explanations.Comparative training curves further reveal faster convergence and lower variance across seeds.Efficiency analysis shows less than 3%overhead in wall-clock training time and no additional inference cost,making the approach practical for realworld deployment.This study demonstrates that a simple task-structured curriculum can significantly improve the effectiveness of knowledge distillation.By separating and sequencing objectives,TSCL achieves a better balance between accuracy,stability,and explanation quality.The framework generalizes across domains,including medical NLI,and offers a principled recipe for future applications in multimodal reasoning and reinforcement learning.展开更多
基金supported by the European Research Council(ERC)under Grant Agreement No.951424(Water-Futures)by the Republic of Cyprus through the Deputy Ministry of Research,Innovation and Digital Policy.
文摘In today's connected world,the generation of massive streaming data across diverse domains has become commonplace.In the presence of concept drift,class imbalance,label scarcity,and new class emergence,these challenges jointly degrade representation stability,bias learning toward outdated distributions,and reduce the resilience and reliability of detection in dynamic environments.This paper proposes a streaming classincremental learning(SCIL)framework to address these issues.The SCIL framework integrates an autoencoder(AE)with a multi-layer perceptron for multi-class prediction,employs a dual-loss strategy(classification and reconstruction)for prediction and new class detection,uses corrected pseudo-labels for online training,manages classes with queues,and applies oversampling to handle imbalance.The rationale behind the method's structure is elucidated through ablation studies,and a comprehensive experimental evaluation is performed using both real-world and synthetic datasets that feature class imbalance,incremental classes,and concept drifts.Our results demonstrate that SCIL outperforms strong baselines and state-of-the-art methods.In line with our commitment to Open Science,we make our code and datasets available to the community.
基金supported by the Gansu Provincial Natural Science Foundation(grant number 25JRRA074)the Gansu Provincial Key R&D Science and Technology Program(grant number 24YFGA060)the National Natural Science Foundation of China(grant number 62161019).
文摘Modern intelligent systems,such as autonomous vehicles and face recognition,must continuously adapt to new scenarios while preserving their ability to handle previously encountered situations.However,when neural networks learn new classes sequentially,they suffer from catastrophic forgetting—the tendency to lose knowledge of earlier classes.This challenge,which lies at the core of class-incremental learning,severely limits the deployment of continual learning systems in real-world applications with streaming data.Existing approaches,including rehearsalbased methods and knowledge distillation techniques,have attempted to address this issue but often struggle to effectively preserve decision boundaries and discriminative features under limited memory constraints.To overcome these limitations,we propose a support vector-guided framework for class-incremental learning.The framework integrates an enhanced feature extractor with a Support Vector Machine classifier,which generates boundary-critical support vectors to guide both replay and distillation.Building on this architecture,we design a joint feature retention strategy that combines boundary proximity with feature diversity,and a Support Vector Distillation Loss that enforces dual alignment in decision and semantic spaces.In addition,triple attention modules are incorporated into the feature extractor to enhance representation power.Extensive experiments on CIFAR-100 and Tiny-ImageNet demonstrate effective improvements.On CIFAR-100 and Tiny-ImageNet with 5 tasks,our method achieves 71.68%and 58.61%average accuracy,outperforming strong baselines by 3.34%and 2.05%.These advantages are consistently observed across different task splits,highlighting the robustness and generalization of the proposed approach.Beyond benchmark evaluations,the framework also shows potential in few-shot and resource-constrained applications such as edge computing and mobile robotics.
基金supported by the National Natural Science Foundation of China(Grants 62406071,U21A20471)Fujian Provincial Natural Science Foundation(Grant 2022J05135).
文摘Continual learning aims to empower a model to learn new tasks continuously while reducing forgetting to retain previously learnt knowledge.In the context of receiving streaming data that are not constrained by the independent and identically distributed(IID)assumption,continual learning efficiently transforms and leverages previously learnt knowledge through various methodologies and completes the learning of new tasks.The generalisation performance and learning efficiency of the model are enhanced in a sequence of tasks.However,the class imbalance in continual learning scenarios critically undermines model performance.In particular,in the class-incremental scenario,the class imbalance results in a bias towards new task classes while degrading the performance on previous learnt classes,leading to catastrophic forgetting.In this paper,a novel method based on balanced contrast is proposed to solve the class-incremental continual learning.The method utilises gradient balancing to mitigate the impact of class imbalance in the class-incremental scenario.The method leverages contrastive learning and gradient modifications to facilitate balanced processing of data across different classes in continual learning.The method proposed in this paper surpasses the existing baseline approaches in the class-incremental learning scenario on standard image datasets such as CIFAR-100,CIFAR-10 and mini-ImageNet.The research results reveal that the proposed method effectively mitigates catastrophic forgetting of previously learnt classes,markedly improving the efficacy of continual learning and offering a powerful solution for further advancing continual learning performance.
基金support from the Strategic Priority Research Program of the Chinese Academy of Sciences under Grant No.XDA27000000.
文摘Deep Convolution Neural Networks(DCNNs)can capture discriminative features from large datasets.However,how to incrementally learn new samples without forgetting old ones and recognize novel classes that arise in the dynamically changing world,e.g.,classifying newly discovered fish species,remains an open problem.We address an even more challenging and realistic setting of this problem where new class samples are insufficient,i.e.,Few-Shot Class-Incremental Learning(FSCIL).Current FSCIL methods augment the training data to alleviate the overfitting of novel classes.By contrast,we propose Filter Bank Networks(FBNs)that augment the learnable filters to capture fine-detailed features for adapting to future new classes.In the forward pass,FBNs augment each convolutional filter to a virtual filter bank containing the canonical one,i.e.,itself,and multiple transformed versions.During back-propagation,FBNs explicitly stimulate fine-detailed features to emerge and collectively align all gradients of each filter bank to learn the canonical one.FBNs capture pattern variants that do not yet exist in the pretraining session,thus making it easy to incorporate new classes in the incremental learning phase.Moreover,FBNs introduce model-level prior knowledge to efficiently utilize the limited few-shot data.Extensive experiments on MNIST,CIFAR100,CUB200,andMini-ImageNet datasets show that FBNs consistently outperformthe baseline by a significantmargin,reporting new state-of-the-art FSCIL results.In addition,we contribute a challenging FSCIL benchmark,Fishshot1K,which contains 8261 underwater images covering 1000 ocean fish species.The code is included in the supplementary materials.
基金supported by the National Key Research and Development Program of China under Grant No.2023YFB2703700the National Natural Science Foundation of China under Grant No.62472410.
文摘Federated learning(FL)enables collaborative model training among participants while guaranteeing the privacy of raw data.Mainstream FL methodologies overlook the dynamic nature of real-world data,particularly its tendency to grow in volume and diversify in classes over time.This oversight results in FL methods suffering from catastrophic forgetting,where the trained models inadvertently discard previously learned information upon assimilating new data.In response to this challenge,we propose a novel federated class-incremental learning(FCIL)method,named Federated Classincremental Learning with New-Class Augmented Self-Distillation(FedCLASS).The core of FedCLASS is to enrich the class scores of historical models with new class scores predicted by current models and utilize the combined knowledge for self-distillation,enabling a more sufficient and precise knowledge transfer from historical models to current models.Theoretical analyses demonstrate that FedCLASS stands on reliable foundations,considering the scores of old classes predicted by historical models as conditional probabilities in the absence of new classes,and the scores of new classes predicted by current models as the conditional probabilities of class scores derived from historical models.Empirical experiments demonstrate the superiority of FedCLASS over four baseline algorithms in reducing average forgetting rate and boosting global accuracy.
基金supported by the National Natural Science Foundation of China(62371465)Taishan Scholar Project of Shandong Province(ts201511020)。
文摘In wireless sensor networks,ensuring communication security via specific emitter identification(SEI)is crucial.However,existing SEI methods are limited to closed-set scenarios and lack the ability to detect unknown devices and perform classincremental training.This study proposes a class-incremental open-set SEI approach.The open-set SEI model calculates radiofrequency fingerprints(RFFs)prototypes for known signals and employs a self-attention mechanism to enhance their discriminability.Detection thresholds are set through Gaussian fitting for each class.For class-incremental learning,the algorithm freezes the parameters of the previously trained model to initialize the new model.It designs specific losses:the RFFs extraction distribution difference loss and the prototype transformation distribution difference loss,which force the new model to retain old knowledge while learning new knowledge.The training loss enables learning of new class RFFs.Experimental results demonstrate that the open-set SEI model achieves state-of-theart performance and strong noise robustness.Moreover,the class-incremental learning algorithm effectively enables the model to retain old device RFFs knowledge,acquire new device RFFs knowledge,and detect unknown devices simultaneously.
文摘1 Introduction Current continual learning methods[1–4]can utilize labeled data to alleviate catastrophic forgetting effectively.However,obtaining labeled samples can be difficult and tedious as it may require expert knowledge.In many practical application scenarios,labeled and unlabeled samples exist simultaneously,with more unlabeled than labeled samples in streaming data[5,6].Unfortunately,existing class-incremental learning methods face limitations in effectively utilizing unlabeled data,thereby impeding their performance in incremental learning scenarios.
基金supported in part by the National Natura Science Foundation of China(U2013602,61876181,51521003)the Nationa Key R&D Program of China(2020YFB13134)+2 种基金Shenzhen Science and Technology Research and Development Foundation(JCYJ20190813171009236)Beijing Nova Program of Science and Technology(Z191100001119043)the Youth Innovation Promotion Association,Chinese Academy of Sciences。
文摘Continual learning(CL)studies the problem of learning to accumulate knowledge over time from a stream of data.A crucial challenge is that neural networks suffer from performance degradation on previously seen data,known as catastrophic forgetting,due to allowing parameter sharing.In this work,we consider a more practical online class-incremental CL setting,where the model learns new samples in an online manner and may continuously experience new classes.Moreover,prior knowledge is unavailable during training and evaluation.Existing works usually explore sample usages from a single dimension,which ignores a lot of valuable supervisory information.To better tackle the setting,we propose a novel replay-based CL method,which leverages multi-level representations produced by the intermediate process of training samples for replay and strengthens supervision to consolidate previous knowledge.Specifically,besides the previous raw samples,we store the corresponding logits and features in the memory.Furthermore,to imitate the prediction of the past model,we construct extra constraints by leveraging multi-level information stored in the memory.With the same number of samples for replay,our method can use more past knowledge to prevent interference.We conduct extensive evaluations on several popular CL datasets,and experiments show that our method consistently outperforms state-of-the-art methods with various sizes of episodic memory.We further provide a detailed analysis of these results and demonstrate that our method is more viable in practical scenarios.
基金Guangzhou Metro Scientific Research Project(No.JT204-100111-23001)Chongqing Municipal Special Project for Technological Innovation and Application Development(No.CSTB2022TIAD-KPX0101)Science and Technology Research and Development Program of China State Railway Group Co.,Ltd.(No.N2023G045)。
文摘The uplift resistance of the soil overlying shield tunnels significantly impacts their anti-floating stability.However,research on uplift resistance concerning special-shaped shield tunnels is limited.This study combines numerical simulation with machine learning techniques to explore this issue.It presents a summary of special-shaped tunnel geometries and introduces a shape coefficient.Through the finite element software,Plaxis3D,the study simulates six key parameters—shape coefficient,burial depth ratio,tunnel’s longest horizontal length,internal friction angle,cohesion,and soil submerged bulk density—that impact uplift resistance across different conditions.Employing XGBoost and ANN methods,the feature importance of each parameter was analyzed based on the numerical simulation results.The findings demonstrate that a tunnel shape more closely resembling a circle leads to reduced uplift resistance in the overlying soil,whereas other parameters exhibit the contrary effects.Furthermore,the study reveals a diminishing trend in the feature importance of buried depth ratio,internal friction angle,tunnel longest horizontal length,cohesion,soil submerged bulk density,and shape coefficient in influencing uplift resistance.
基金supported in part by the National Key Research and Development Program of China under Grant 2021YFF1200600in part by the National Natural Science Foundation of China under Grant 62301523.
文摘Hand gesture recognition(HGR)plays a vital role in human-computer interaction.The integration of high-density surface electromyography(HD-sEMG)and deep neural networks(DNNs)has significantly improved the robustness and accuracy of HGR systems.These methods are typically effective for a fixed set of trained gestures.However,the need for new gesture classes over time poses a challenge.Introducing new classes to DNNs can lead to a substantial decrease in accuracy for previously learned tasks,a phenomenon known as“catastrophic forgetting,”especially when the training data for earlier tasks is not retained and retrained.This issue is exacerbated in embedded devices with limited storage,which struggle to store the large-scale data of HD-sEMG.Classincremental learning(CIL)is an effective method to reduce catastrophic forgetting.However,existing CIL methods for HGR rarely focus on reducing memory load.To address this,we propose a memory-friendly CIL method for HGR using HD-sEMG.Our approach includes a lightweight convolutional neural network,named SeparaNet,for feature representation learning,coupled with a nearest-mean-of-exemplars classifier for classifi-cation.We introduce a priority exemplar selection algorithm inspired by the herding effect to maintain a manageable set of exemplars during training.Furthermore,a task-equal-weight exemplar sampling strategy is proposed to effectively reduce memory load while preserving high recognition performance.Experimental results on two datasets demonstrate that our method significantly reduces the number of retained exemplars to only a quarter of that required by other CIL methods,accounting for less than 5%of the total samples,while still achieving comparable average accuracy.
基金supported by the National Natural Science Foundation of China (42505149,41925023,U2342223,42105069,and 91744208)the China Postdoctoral Science Foundation (2025M770303)+1 种基金the Fundamental Research Funds for the Central Universities (14380230)the Jiangsu Funding Program for Excellent Postdoctoral Talent,and Jiangsu Collaborative Innovation Center of Climate Change。
文摘Countries around the world have been making efforts to reduce pollutant emissions. However, the response of global black carbon(BC) aging to emission changes remains unclear. Using the Community Atmosphere Model version 6 with a machine-learning-integrated four-mode version of the Modal Aerosol Module, we quantify global BC aging responses to emission reductions for 2011–2018 and for 2050 and 2100 under carbon neutrality. During 2011–18, global trends in BC aging degree(mass ratio of coatings to BC, R_(BC)) exhibited marked regional disparities, with a significant increase in China(5.4% yr^(-1)), which contrasts with minimal changes in the USA, Europe, and India. The divergence is attributed to opposing trends in secondary organic aerosol(SOA) and sulfate coatings, driven by regional changes in the emission ratios of corresponding coating precursors to BC(volatile organic compounds-VOCs/BC and SO_(2)/BC). Projections under carbon neutrality reveal that R_(BC) will increase globally by 47%(118%) in 2050(2100), with strong convergent increases expected across major source regions. The R_(BC) increase, primarily driven by enhanced SOA coatings due to sharper BC reductions relative to VOCs, will enhance the global BC mass absorption cross-section(MAC) by 11%(17%) in 2050(2100).Consequently, although the global BC burden will decline sharply by 60%(76%), the enhanced MAC partially offsets the magnitude of the decline in the BC direct radiative effect, resulting in the moderation of global BC DRE decreases to 88%(92%) of the BC burden reductions in 2050(2100). This study highlights the globally enhanced BC aging and light absorption capacity under carbon neutrality, thereby partly offsetting the impact of BC direct emission reductions on future changes in BC radiative effects globally.
文摘Underwater pipeline inspection plays a vital role in the proactive maintenance and management of critical marine infrastructure and subaquatic systems.However,the inspection of underwater pipelines presents a challenge due to factors such as light scattering,absorption,restricted visibility,and ambient noise.The advancement of deep learning has introduced powerful techniques for processing large amounts of unstructured and imperfect data collected from underwater environments.This study evaluated the efficacy of the You Only Look Once(YOLO)algorithm,a real-time object detection and localization model based on convolutional neural networks,in identifying and classifying various types of pipeline defects in underwater settings.YOLOv8,the latest evolution in the YOLO family,integrates advanced capabilities,such as anchor-free detection,a cross-stage partial network backbone for efficient feature extraction,and a feature pyramid network+path aggregation network neck for robust multi-scale object detection,which make it particularly well-suited for complex underwater environments.Due to the lack of suitable open-access datasets for underwater pipeline defects,a custom dataset was captured using a remotely operated vehicle in a controlled environment.This application has the following assets available for use.Extensive experimentation demonstrated that YOLOv8 X-Large consistently outperformed other models in terms of pipe defect detection and classification and achieved a strong balance between precision and recall in identifying pipeline cracks,rust,corners,defective welds,flanges,tapes,and holes.This research establishes the baseline performance of YOLOv8 for underwater defect detection and showcases its potential to enhance the reliability and efficiency of pipeline inspection tasks in challenging underwater environments.
基金Supported by CAS Basic and Interdisciplinary Frontier Scientific Research Pilot Project(XDB1190300,XDB1190302)Youth Innovation Promotion Association CAS(Y2021056)+1 种基金Joint Fund of the Yulin University and the Dalian National Laboratory for Clean Energy(YLU-DNL Fund 2022007)The special fund for Science and Technology Innovation Teams of Shanxi Province(202304051001007)。
文摘Cyclohexene is an important raw material in the production of nylon.Selective hydrogenation of benzene is a key method for preparing cyclohexene.However,the Ru catalysts used in current industrial processes still face challenges,including high metal usage,high process costs,and low cyclohexene yield.This study utilizes existing literature data combined with machine learning methods to analyze the factors influencing benzene conversion,cyclohexene selectivity,and yield in the benzene hydrogenation to cyclohexene reaction.It constructs predictive models based on XGBoost and Random Forest algorithms.After analysis,it was found that reaction time,Ru content,and space velocity are key factors influencing cyclohexene yield,selectivity,and benzene conversion.Shapley Additive Explanations(SHAP)analysis and feature importance analysis further revealed the contribution of each variable to the reaction outcomes.Additionally,we randomly generated one million variable combinations using the Dirichlet distribution to attempt to predict high-yield catalyst formulations.This paper provides new insights into the application of machine learning in heterogeneous catalysis and offers some reference for further research.
文摘The electrocatalytic reduction of nitric oxide for ammonia synthesis(NORR)is a key green energy conversion technology.Its efficiency relies on high-performance electrocatalysts to enhance both ammonia yield(Y_(NH3))and Faradaic efficiency(F_(NH3)).However,conventional experimental methods for screening high-activity NORR catalysts often entail high resource consumption and time costs.Machine learning combined with SHAP feature analysis was employed to establish a stacked ensemble model that integrates multiple algorithms,to allow for a systematic investigation of the key descriptors governing NORR performance based on an experimental dataset.Evaluation of eight model algorithms revealed that the Stacked-SVR model achieved an R^(2)of 0.9223 and an RMSE of 0.0608 for predicting on the test set,whereas the Stacked-RF model achieved an R^(2)of 0.9042 and an RMSE of 0.0900 for predicting.The stacked ensemble model integrates the strengths of individual algorithms and demonstrates strong NORR prediction performance while avoiding overfitting.SHAP feature analysis results revealed that the Cu content in the catalyst composition has the most significant impact on catalytic performance.Moreover,the combination of the wet chemical reduction synthesis,a carbon fiber(CF)conductive substrate,and HCl electrolyte is more favorable for enhancing catalytic activity.Additionally,moderately lowering the working potential,controlling the electrolyte volume at low to medium levels,reducing catalyst loading,and increasing electrolyte concentration were found to synergistically enhance both and.
文摘Federated Learning(FL)has become a leading decentralized solution that enables multiple clients to train a model in a collaborative environment without directly sharing raw data,making it suitable for privacy-sensitive applications such as healthcare,finance,and smart systems.As the field continues to evolve,the research field has become more complex and scattered,covering different system designs,training methods,and privacy techniques.This survey is organized around the three core challenges:how the data is distributed,how models are synchronized,and how to defend against attacks.It provides a structured and up-to-date review of FL research from 2023 to 2025,offering a unified taxonomy that categorizes works by data distribution(Horizontal FL,Vertical FL,Federated Transfer Learning,and Personalized FL),training synchronization(synchronous and asynchronous FL),optimization strategies,and threat models(data leakage and poisoning attacks).In particular,we summarize the latest contributions in Vertical FL frameworks for secure multi-party learning,communication-efficient Horizontal FL,and domain-adaptive Federated Transfer Learning.Furthermore,we examine synchronization techniques addressing system heterogeneity,including straggler mitigation in synchronous FL and staleness management in asynchronous FL.The survey covers security threats in FL,such as gradient inversion,membership inference,and poisoning attacks,as well as their defense strategies that include privacy-preserving aggregation and anomaly detection.The paper concludes by outlining unresolved issues and highlighting challenges in handling personalized models,scalability,and real-world adoption.
基金supported by a grant(No.CRPG-25-2054)under the Cybersecurity Research and Innovation Pioneers Initiative,provided by the National Cybersecurity Authority(NCA)in the Kingdom of Saudi Arabia.
文摘Split Learning(SL)has been promoted as a promising collaborative machine learning technique designed to address data privacy and resource efficiency.Specifically,neural networks are divided into client and server subnetworks in order to mitigate the exposure of sensitive data and reduce the overhead on client devices,thereby making SL particularly suitable for resource-constrained devices.Although SL prevents the direct transmission of raw data,it does not alleviate entirely the risk of privacy breaches.In fact,the data intermediately transmitted to the server sub-model may include patterns or information that could reveal sensitive data.Moreover,achieving a balance between model utility and data privacy has emerged as a challenging problem.In this article,we propose a novel defense approach that combines:(i)Adversarial learning,and(ii)Network channel pruning.In particular,the proposed adversarial learning approach is specifically designed to reduce the risk of private data exposure while maintaining high performance for the utility task.On the other hand,the suggested channel pruning enables the model to adaptively adjust and reactivate pruned channels while conducting adversarial training.The integration of these two techniques reduces the informativeness of the intermediate data transmitted by the client sub-model,thereby enhancing its robustness against attribute inference attacks without adding significant computational overhead,making it wellsuited for IoT devices,mobile platforms,and Internet of Vehicles(IoV)scenarios.The proposed defense approach was evaluated using EfficientNet-B0,a widely adopted compact model,along with three benchmark datasets.The obtained results showcased its superior defense capability against attribute inference attacks compared to existing state-of-the-art methods.This research’s findings demonstrated the effectiveness of the proposed channel pruning-based adversarial training approach in achieving the intended compromise between utility and privacy within SL frameworks.In fact,the classification accuracy attained by the attackers witnessed a drastic decrease of 70%.
文摘The Internet of Vehicles,or IoV,is expected to lessen pollution,ease traffic,and increase road safety.IoV entities’interconnectedness,however,raises the possibility of cyberattacks,which can have detrimental effects.IoV systems typically send massive volumes of raw data to central servers,which may raise privacy issues.Additionally,model training on IoV devices with limited resources normally leads to slower training times and reduced service quality.We discuss a privacy-preserving Federated Split Learning with Tiny Machine Learning(TinyML)approach,which operates on IoV edge devices without sharing sensitive raw data.Specifically,we focus on integrating split learning(SL)with federated learning(FL)and TinyML models.FL is a decentralisedmachine learning(ML)technique that enables numerous edge devices to train a standard model while retaining data locally collectively.The article intends to thoroughly discuss the architecture and challenges associated with the increasing prevalence of SL in the IoV domain,coupled with FL and TinyML.The approach starts with the IoV learning framework,which includes edge computing,FL,SL,and TinyML,and then proceeds to discuss how these technologies might be integrated.We elucidate the comprehensive operational principles of Federated and split learning by examining and addressingmany challenges.We subsequently examine the integration of SL with FL and various applications of TinyML.Finally,exploring the potential integration of FL and SL with TinyML in the IoV domain is referred to as FSL-TM.It is a superior method for preserving privacy as it conducts model training on individual devices or edge nodes,thereby obviating the necessity for centralised data aggregation,which presents considerable privacy threats.The insights provided aim to help both researchers and practitioners understand the complicated terrain of FL and SL,hence facilitating advancement in this swiftly progressing domain.
文摘Knowledge distillation has become a standard technique for compressing large language models into efficient student models,but existing methods often struggle to balance prediction accuracy with explanation quality.Recent approaches such as Distilling Step-by-Step(DSbS)introduce explanation supervision,yet they apply it in a uniform manner that may not fully exploit the different learning dynamics of prediction and explanation.In this work,we propose a task-structured curriculum learning(TSCL)framework that structures training into three sequential phases:(i)prediction-only,to establish stable feature representations;(ii)joint prediction-explanation,to align task outputs with rationale generation;and(iii)explanation-only,to refine the quality of rationales.This design provides a simple but effective modification to DSbS,requiring no architectural changes and adding negligible training cost.We justify the phase scheduling with ablation studies and convergence analysis,showing that an initial prediction-heavy stage followed by a balanced joint phase improves both stability and explanation alignment.Extensive experiments on five datasets(e-SNLI,ANLI,CommonsenseQA,SVAMP,and MedNLI)demonstrate that TSCL consistently outperforms strong baselines,achieving gains of+1.7-2.6 points in accuracy and 0.8-1.2 in ROUGE-L,corresponding to relative error reductions of up to 21%.Beyond lexical metrics,human evaluation and ERASERstyle faithfulness diagnostics confirm that TSCL produces more faithful and informative explanations.Comparative training curves further reveal faster convergence and lower variance across seeds.Efficiency analysis shows less than 3%overhead in wall-clock training time and no additional inference cost,making the approach practical for realworld deployment.This study demonstrates that a simple task-structured curriculum can significantly improve the effectiveness of knowledge distillation.By separating and sequencing objectives,TSCL achieves a better balance between accuracy,stability,and explanation quality.The framework generalizes across domains,including medical NLI,and offers a principled recipe for future applications in multimodal reasoning and reinforcement learning.
