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
Underground engineering projects such as deep tunnel excavation often encounter rockburst disasters accompanied by numerous microseismic events.Rapid interpretation of microseismic signals is crucial for the timely id...Underground engineering projects such as deep tunnel excavation often encounter rockburst disasters accompanied by numerous microseismic events.Rapid interpretation of microseismic signals is crucial for the timely identification of rockbursts.However,conventional processing encompasses multi-step workflows,including classification,denoising,picking,locating,and computational analysis,coupled with manual intervention,which collectively compromise the reliability of early warnings.To address these challenges,this study innovatively proposes the“microseismic stethoscope"-a multi-task machine learning and deep learning model designed for the automated processing of massive microseismic signals.This model efficiently extracts three key parameters that are necessary for recognizing rockburst disasters:rupture location,microseismic energy,and moment magnitude.Specifically,the model extracts raw waveform features from three dedicated sub-networks:a classifier for source zone classification,and two regressors for microseismic energy and moment magnitude estimation.This model demonstrates superior efficiency compared to traditional processing and semi-automated processing,reducing per-event processing time from 0.71 s to 0.49 s to merely 0.036 s.It concurrently achieves 98%accuracy in source zone classification,with microseismic energy and moment magnitude estimation errors of 0.13 and 0.05,respectively.This model has been well applied and validated in the Daxiagu Tunnel case in Sichuan,China.The application results indicate that the model is as accurate as traditional methods in determining source parameters,and thus can be used to identify potential geomechanical processes of rockburst disasters.By enhancing the signal processing reliability of microseismic events,the proposed model in this study presents a significant advancement in the identification of rockburst disasters.展开更多
Surface properties of crystals are critical in many fields,including electrochemistry and photoelectronics,the efficient prediction of which can expedite the design and optimization of catalysts,batteries,alloys etc.H...Surface properties of crystals are critical in many fields,including electrochemistry and photoelectronics,the efficient prediction of which can expedite the design and optimization of catalysts,batteries,alloys etc.However,we are still far from realizing this vision due to the rarity of surface property-related databases,especially for multicomponent compounds,due to the large sample spaces and limited computing resources.In this work,we present a surface emphasized multi-task crystal graph convolutional neural network(SEM-CGCNN)to predict multiple surface properties simultaneously from crystal structures.The model is evaluated on a dataset of 3526 surface energies and work functions of binary magnesium intermetallics obtained through first-principles calculations,and obvious improvements are observed both in efficiency and accuracy over the original CGCNN model.By transferring the pre-trained model to the datasets of pure metals and other intermetallics,the fine-tuned SEM-CGCNN outperforms learning from scratch and can be further applied to other surface properties and materials systems.This study could be a paradigm for the end-to-end mapping of atomic structures to anisotropic surface properties of crystals,which provides an efficient framework to understand and screen materials with desired surface characteristics.展开更多
Reconfigurable intelligent surface(RIS)have been cast as a promising alternative to alleviate blockage vulnerability and enhance coverage capability for terahertz(THz)communications.Owing to large-scale array elements...Reconfigurable intelligent surface(RIS)have been cast as a promising alternative to alleviate blockage vulnerability and enhance coverage capability for terahertz(THz)communications.Owing to large-scale array elements at transceivers and RIS,the codebook based beamforming can be utilized in a computationally efficient manner.However,the codeword selection for analog beamforming is an intractable combinatorial optimization(CO)problem.To this end,by taking the CO problem as a classification problem,a multi-task learning based analog beam selection(MTL-ABS)framework is developed to implement cooperative beam selection concurrently at transceivers and RIS.In addition,residual network and self-attention mechanism are used to combat the network degradation and mine intrinsic THz channel features.Finally,the network convergence is analyzed from a blockwise perspective,and numerical results demonstrate that the MTL-ABS framework greatly decreases the beam selection overhead and achieves near optimal sum-rate compared with heuristic search based counterparts.展开更多
Satellite communication technology has emerged as a key solution to address the challenges of data transmission in remote areas.By overcoming the limitations of traditional terrestrial communication networks,it enable...Satellite communication technology has emerged as a key solution to address the challenges of data transmission in remote areas.By overcoming the limitations of traditional terrestrial communication networks,it enables long-distance data transmission anytime and anywhere,ensuring the timely and accurate delivery of water level data,which is particularly crucial for fishway water level monitoring.To enhance the effectiveness of fishway water level monitoring,this study proposes a multi-task learning model,AS-SOMTF,designed for real-time and comprehensive prediction.The model integrates auxiliary sequences with primary input sequences to capture complex relationships and dependencies,thereby improving representational capacity.In addition,a novel timeseries embedding algorithm,AS-SOM,is introduced,which combines generative inference and pooling operations to optimize prediction efficiency for long sequences.This innovation not only ensures the timely transmission of water level data but also enhances the accuracy of real-time monitoring.Compared with traditional models such as Transformer and Long Short-Term Memory(LSTM)networks,the proposed model achieves improvements of 3.8%and 1.4%in prediction accuracy,respectively.These advancements provide more precise technical support for water level forecasting and resource management in the Diqing Tibetan Autonomous Prefecture of the Lancang River,contributing to ecosystem protection and improved operational safety.展开更多
Deep neural networks(DNNs)are effective in solving both forward and inverse problems for nonlinear partial differential equations(PDEs).However,conventional DNNs are not effective in handling problems such as delay di...Deep neural networks(DNNs)are effective in solving both forward and inverse problems for nonlinear partial differential equations(PDEs).However,conventional DNNs are not effective in handling problems such as delay differential equations(DDEs)and delay integrodifferential equations(DIDEs)with constant delays,primarily due to their low regularity at delayinduced breaking points.In this paper,a DNN method that combines multi-task learning(MTL)which is proposed to solve both the forward and inverse problems of DIDEs.The core idea of this approach is to divide the original equation into multiple tasks based on the delay,using auxiliary outputs to represent the integral terms,followed by the use of MTL to seamlessly incorporate the properties at the breaking points into the loss function.Furthermore,given the increased training dificulty associated with multiple tasks and outputs,we employ a sequential training scheme to reduce training complexity and provide reference solutions for subsequent tasks.This approach significantly enhances the approximation accuracy of solving DIDEs with DNNs,as demonstrated by comparisons with traditional DNN methods.We validate the effectiveness of this method through several numerical experiments,test various parameter sharing structures in MTL and compare the testing results of these structures.Finally,this method is implemented to solve the inverse problem of nonlinear DIDE and the results show that the unknown parameters of DIDE can be discovered with sparse or noisy data.展开更多
Cooperative multi-agent reinforcement learning(MARL)is a key technology for enabling cooperation in complex multi-agent systems.It has achieved remarkable progress in areas such as gaming,autonomous driving,and multi-...Cooperative multi-agent reinforcement learning(MARL)is a key technology for enabling cooperation in complex multi-agent systems.It has achieved remarkable progress in areas such as gaming,autonomous driving,and multi-robot control.Empowering cooperative MARL with multi-task decision-making capabilities is expected to further broaden its application scope.In multi-task scenarios,cooperative MARL algorithms need to address 3 types of multi-task problems:reward-related multi-task,arising from different reward functions;multi-domain multi-task,caused by differences in state and action spaces,state transition functions;and scalability-related multi-task,resulting from the dynamic variation in the number of agents.Most existing studies focus on scalability-related multitask problems.However,with the increasing integration between large language models(LLMs)and multi-agent systems,a growing number of LLM-based multi-agent systems have emerged,enabling more complex multi-task cooperation.This paper provides a comprehensive review of the latest advances in this field.By combining multi-task reinforcement learning with cooperative MARL,we categorize and analyze the 3 major types of multi-task problems under multi-agent settings,offering more fine-grained classifications and summarizing key insights for each.In addition,we summarize commonly used benchmarks and discuss future directions of research in this area,which hold promise for further enhancing the multi-task cooperation capabilities of multi-agent systems and expanding their practical applications in the real world.展开更多
The accurate prediction of drug absorption,distribution,metabolism,excretion,and toxicity(ADMET)properties represents a crucial step in early drug development for reducing failure risk.Current deep learning approaches...The accurate prediction of drug absorption,distribution,metabolism,excretion,and toxicity(ADMET)properties represents a crucial step in early drug development for reducing failure risk.Current deep learning approaches face challenges with data sparsity and information loss due to single-molecule representation limitations and isolated predictive tasks.This research proposes molecular properties prediction with parallel-view and collaborative learning(MolP-PC),a multi-view fusion and multi-task deep learning framework that integrates 1D molecular fingerprints(MFs),2D molecular graphs,and 3D geometric representations,incorporating an attention-gated fusion mechanism and multi-task adaptive learning strategy for precise ADMET property predictions.Experimental results demonstrate that MolP-PC achieves optimal performance in 27 of 54 tasks,with its multi-task learning(MTL)mechanism significantly enhancing predictive performance on small-scale datasets and surpassing single-task models in 41 of 54 tasks.Additional ablation studies and interpretability analyses confirm the significance of multi-view fusion in capturing multi-dimensional molecular information and enhancing model generalization.A case study examining the anticancer compound Oroxylin A demonstrates MolP-PC’s effective generalization in predicting key pharmacokinetic parameters such as half-life(T0.5)and clearance(CL),indicating its practical utility in drug modeling.However,the model exhibits a tendency to underestimate volume of distribution(VD),indicating potential for improvement in analyzing compounds with high tissue distribution.This study presents an efficient and interpretable approach for ADMET property prediction,establishing a novel framework for molecular optimization and risk assessment in drug development.展开更多
Predicting the health status of stroke patients at different stages of the disease is a critical clinical task.The onset and development of stroke are affected by an array of factors,encompassing genetic predispositio...Predicting the health status of stroke patients at different stages of the disease is a critical clinical task.The onset and development of stroke are affected by an array of factors,encompassing genetic predisposition,environmental exposure,unhealthy lifestyle habits,and existing medical conditions.Although existing machine learning-based methods for predicting stroke patients’health status have made significant progress,limitations remain in terms of prediction accuracy,model explainability,and system optimization.This paper proposes a multi-task learning approach based on Explainable Artificial Intelligence(XAI)for predicting the health status of stroke patients.First,we design a comprehensive multi-task learning framework that utilizes the task correlation of predicting various health status indicators in patients,enabling the parallel prediction of multiple health indicators.Second,we develop a multi-task Area Under Curve(AUC)optimization algorithm based on adaptive low-rank representation,which removes irrelevant information from the model structure to enhance the performance of multi-task AUC optimization.Additionally,the model’s explainability is analyzed through the stability analysis of SHAP values.Experimental results demonstrate that our approach outperforms comparison algorithms in key prognostic metrics F1 score and Efficiency.展开更多
Aerosol optical depth(AOD)and fine particulate matter with a diameter of less than or equal to 2.5μm(PM_(2.5))play crucial roles in air quality,human health,and climate change.However,the complex correlation of AOD–...Aerosol optical depth(AOD)and fine particulate matter with a diameter of less than or equal to 2.5μm(PM_(2.5))play crucial roles in air quality,human health,and climate change.However,the complex correlation of AOD–PM_(2.5)and the limitations of existing algorithms pose a significant challenge in realizing the accurate joint retrieval of these two parameters at the same location.On this point,a multi-task learning(MTL)model,which enables the joint retrieval of PM_(2.5)concentration and AOD,is proposed and applied on the top-of-the-atmosphere reflectance data gathered by the Fengyun-4A Advanced Geosynchronous Radiation Imager(FY-4A AGRI),and compared to that of two single-task learning models—namely,Random Forest(RF)and Deep Neural Network(DNN).Specifically,MTL achieves a coefficient of determination(R^(2))of 0.88 and a root-mean-square error(RMSE)of 0.10 in AOD retrieval.In comparison to RF,the R^(2)increases by 0.04,the RMSE decreases by 0.02,and the percentage of retrieval results falling within the expected error range(Within-EE)rises by 5.55%.The R^(2)and RMSE of PM_(2.5)retrieval by MTL are 0.84 and 13.76μg m~(-3)respectively.Compared with RF,the R^(2)increases by 0.06,the RMSE decreases by 4.55μg m~(-3),and the Within-EE increases by 7.28%.Additionally,compared to DNN,MTL shows an increase of 0.01 in R^(2)and a decrease of 0.02 in RMSE in AOD retrieval,with a corresponding increase of 2.89%in Within-EE.For PM_(2.5)retrieval,MTL exhibits an increase of 0.05 in R^(2),a decrease of 1.76μg m~(-3)in RMSE,and an increase of 6.83%in Within-EE.The evaluation suggests that MTL is able to provide simultaneously improved AOD and PM_(2.5)retrievals,demonstrating a significant advantage in efficiently capturing the spatial distribution of PM_(2.5)concentration and AOD.展开更多
Accurate short-term forecast of offshore wind fields is still challenging for numerical weather prediction models.Based on three years of 48-hour forecast data from the European Centre for Medium-Range Weather Forecas...Accurate short-term forecast of offshore wind fields is still challenging for numerical weather prediction models.Based on three years of 48-hour forecast data from the European Centre for Medium-Range Weather Forecasts Integrated Forecasting System global model(ECMWF-IFS)over 14 offshore weather stations along the coast of Shandong Province,this study introduces a multi-task learning(MTL)model(TabNet-MTL),which significantly improves the forecast bias of near-surface wind direction and speed simultaneously.TabNet-MTL adopts the feature engineering method,utilizes mean square error as the loss function,and employs the 5-fold cross validation method to ensure the generalization ability of the trained model.It demonstrates superior skills in wind field correction across different forecast lead times over all stations compared to its single-task version(TabNet-STL)and three other popular single-task learning models(Random Forest,LightGBM,and XGBoost).Results show that it significantly reduces root mean square error of the ECMWF-IFS wind speed forecast from 2.20 to 1.25 m s−1,and increases the forecast accuracy of wind direction from 50%to 65%.As an explainable deep learning model,the weather stations and long-term temporal statistics of near-surface wind speed are identified as the most influential variables for TabNet-MTL in constructing its feature engineering.展开更多
Staple crops are the cornerstone of the food supply but are frequently threatened by plant diseases.Effective disease management,including disease identification and severity assessment,helps to better address these c...Staple crops are the cornerstone of the food supply but are frequently threatened by plant diseases.Effective disease management,including disease identification and severity assessment,helps to better address these challenges.Currently,methods for disease severity assessment typically rely on calculating the area proportion of disease segmentation regions or using classification networks for severity assessment.However,these methods require large amounts of labeled data and fail to quantify lesion proportions when using classification networks,leading to inaccurate evaluations.To address these issues,we propose an automated framework for disease severity assessment that combines multi-task learning and knowledge-driven large-model segmentation techniques.This framework includes an image information processor,a lesion and leaf segmentation module,and a disease severity assessment module.First,the image information processor utilizes a multi-task learning strategy to analyze input images comprehensively,ensuring a deep understanding of disease characteristics.Second,the lesion and leaf segmentation module employ prompt-driven large-model technology to accurately segment diseased areas and entire leaves,providing detailed visual analysis.Finally,the disease severity assessment module objectively evaluates the severity of the disease based on professional grading standards by calculating lesion area proportions.Additionally,we have developed a comprehensive database of diseased leaf images from major crops,including several task-specific datasets.Experimental results demonstrate that our framework can accurately identify and assess the types and severity of crop diseases,even without extensive labeled data.Codes and data are available at http://dkp-ads.samlab.cn/.展开更多
As the group-buying model shows significant progress in attracting new users,enhancing user engagement,and increasing platform profitability,providing personalized recommendations for group-buying users has emerged as...As the group-buying model shows significant progress in attracting new users,enhancing user engagement,and increasing platform profitability,providing personalized recommendations for group-buying users has emerged as a new challenge in the field of recommendation systems.This paper introduces a group-buying recommendation model based on multi-head attention mechanisms and multi-task learning,termed the Multi-head Attention Mechanisms and Multi-task Learning Group-Buying Recommendation(MAMGBR)model,specifically designed to optimize group-buying recommendations on e-commerce platforms.The core dataset of this study comes from the Chinese maternal and infant e-commerce platform“Beibei,”encompassing approximately 430,000 successful groupbuying actions and over 120,000 users.Themodel focuses on twomain tasks:recommending items for group organizers(Task Ⅰ)and recommending participants for a given group-buying event(Task Ⅱ).In model evaluation,MAMGBR achieves an MRR@10 of 0.7696 for Task I,marking a 20.23%improvement over baseline models.Furthermore,in Task II,where complex interaction patterns prevail,MAMGBR utilizes auxiliary loss functions to effectively model the multifaceted roles of users,items,and participants,leading to a 24.08%increase in MRR@100 under a 1:99 sample ratio.Experimental results show that compared to benchmark models,such as NGCF and EATNN,MAMGBR’s integration ofmulti-head attentionmechanisms,expert networks,and gating mechanisms enables more accurate modeling of user preferences and social associations within group-buying scenarios,significantly enhancing recommendation accuracy and platform group-buying success rates.展开更多
Accurate forecasting of blast furnace gas(BFG)production is an essential prerequisite for reasonable energy scheduling and management to reduce carbon emissions.Coupling forecasting between BFG generation and consumpt...Accurate forecasting of blast furnace gas(BFG)production is an essential prerequisite for reasonable energy scheduling and management to reduce carbon emissions.Coupling forecasting between BFG generation and consumption dynamics was taken as the research object.A multi-task learning(MTL)method for BFG forecasting was proposed,which integrated a coupling correlation coefficient(CCC)and an inverted transformer structure.The CCC method could enhance key information extraction by establishing relationships between multiple prediction targets and relevant factors,while MTL effectively captured the inherent correlations between BFG generation and consumption.Finally,a real-world case study was conducted to compare the proposed model with four benchmark models.Results indicated significant reductions in average mean absolute percentage error by 33.37%,achieving 1.92%,with a computational time of 76 s.The sensitivity analysis of hyperparameters such as learning rate,batch size,and units of the long short-term memory layer highlights the importance of hyperparameter tuning.展开更多
With focus now placed on the learner, more attention is given to his learning style, multiple intelligence and developing learning strategies to enable him to make sense of and use of the target language appropriately...With focus now placed on the learner, more attention is given to his learning style, multiple intelligence and developing learning strategies to enable him to make sense of and use of the target language appropriately in varied contexts and with different uses of the language. To attain this, the teacher is tasked with designing, monitoring and processing language learning activities for students to carry out and in the process learn by doing and reflecting on the learning process they went through as they interacted socially with each other. This paper describes a task named"The Fishbowl Technique"and found to be effective in large ESL classes in the secondary level in the Philippines.展开更多
The global nuclear mass based on the macroscopic-microscopic model was studied by applying a newly designed multi-task learning artificial neural network(MTL-ANN). First, the reported nuclear binding energies of 2095 ...The global nuclear mass based on the macroscopic-microscopic model was studied by applying a newly designed multi-task learning artificial neural network(MTL-ANN). First, the reported nuclear binding energies of 2095 nuclei(Z ≥ 8, N ≥ 8) released in the latest Atomic Mass Evaluation AME2020 and the deviations between the fitting result of the liquid drop model(LDM)and data from AME2020 for each nucleus were obtained.To compensate for the deviations and investigate the possible ignored physics in the LDM, the MTL-ANN method was introduced in the model. Compared to the single-task learning(STL) method, this new network has a powerful ability to simultaneously learn multi-nuclear properties,such as the binding energies and single neutron and proton separation energies. Moreover, it is highly effective in reducing the risk of overfitting and achieving better predictions. Consequently, good predictions can be obtained using this nuclear mass model for both the training and validation datasets and for the testing dataset. In detail, the global root mean square(RMS) of the binding energy is effectively reduced from approximately 2.4 MeV of LDM to the current 0.2 MeV, and the RMS of Sn, Spcan also reach approximately 0.2 MeV. Moreover, compared to STL, for the training and validation sets, 3-9% improvement can be achieved with the binding energy, and 20-30% improvement for S_(n), S_(p);for the testing sets, the reduction in deviations can even reach 30-40%, which significantly illustrates the advantage of the current MTL.展开更多
Recent studies for computer vision and deep learning-based,post-earthquake inspections on RC structures mainly perform well for specific tasks,while the trained models must be fine-tuned and re-trained when facing new...Recent studies for computer vision and deep learning-based,post-earthquake inspections on RC structures mainly perform well for specific tasks,while the trained models must be fine-tuned and re-trained when facing new tasks and datasets,which is inevitably time-consuming.This study proposes a multi-task learning approach that simultaneously accomplishes the semantic segmentation of seven-type structural components,three-type seismic damage,and four-type deterioration states.The proposed method contains a CNN-based encoder-decoder backbone subnetwork with skip-connection modules and a multi-head,task-specific recognition subnetwork.The backbone subnetwork is designed to extract multi-level features of post-earthquake RC structures.The multi-head,task-specific recognition subnetwork consists of three individual self-attention pipelines,each of which utilizes extracted multi-level features from the backbone network as a mutual guidance for the individual segmentation task.A synthetical loss function is designed with real-time adaptive coefficients to balance multi-task losses and focus on the most unstably fluctuating one.Ablation experiments and comparative studies are further conducted to demonstrate their effectiveness and necessity.The results show that the proposed method can simultaneously recognize different structural components,seismic damage,and deterioration states,and that the overall performance of the three-task learning models gains general improvement when compared to all single-task and dual-task models.展开更多
Convective storms and lightning are among the most important weather phenomena that are challenging to forecast.In this study,a novel multi-task learning(MTL)encoder-decoder U-net neural network was developed to forec...Convective storms and lightning are among the most important weather phenomena that are challenging to forecast.In this study,a novel multi-task learning(MTL)encoder-decoder U-net neural network was developed to forecast convective storms and lightning with lead times for up to 90 min,using GOES-16 geostationary satellite infrared brightness temperatures(IRBTs),lightning flashes from Geostationary Lightning Mapper(GLM),and vertically integrated liquid(VIL)from Next Generation Weather Radar(NEXRAD).To cope with the heavily skewed distribution of lightning data,a spatiotemporal exponent-weighted loss function and log-transformed lightning normalization approach were developed.The effects of MTL,single-task learning(STL),and IRBTs as auxiliary input features on convection and lightning nowcasting were investigated.The results showed that normalizing the heavily skew-distributed lightning data along with a log-transformation dramatically outperforms the min-max normalization method for nowcasting an intense lightning event.The MTL model significantly outperformed the STL model for both lightning nowcasting and VIL nowcasting,particularly for intense lightning events.The MTL also helped delay the lightning forecast performance decay with the lead times.Furthermore,incorporating satellite IRBTs as auxiliary input features substantially improved lightning nowcasting,but produced little difference in VIL forecasting.Finally,the MTL model performed better for forecasting both lightning and the VIL of organized convective storms than for isolated cells.展开更多
There is a growing amount of data uploaded to the internet every day and it is important to understand the volume of those data to find a better scheme to process them.However,the volume of internet data is beyond the...There is a growing amount of data uploaded to the internet every day and it is important to understand the volume of those data to find a better scheme to process them.However,the volume of internet data is beyond the processing capabilities of the current internet infrastructure.Therefore,engineering works using technology to organize and analyze information and extract useful information are interesting in both industry and academia.The goal of this paper is to explore the entity relationship based on deep learning,introduce semantic knowledge by using the prepared language model,develop an advanced entity relationship information extraction method by combining Robustly Optimized BERT Approach(RoBERTa)and multi-task learning,and combine the intelligent characters in the field of linguistic,called Robustly Optimized BERT Approach+Multi-Task Learning(RoBERTa+MTL).To improve the effectiveness of model interaction,multi-task teaching is used to implement the observation information of auxiliary tasks.Experimental results show that our method has achieved an accuracy of 88.95 entity relationship extraction,and a further it has achieved 86.35%of accuracy after being combined with multi-task learning.展开更多
Buiding data-driven models using machine learning methods has gradually become a common approach for studying reservoir parameters.Among these methods,deep learning methods are highly effective.From the perspective of...Buiding data-driven models using machine learning methods has gradually become a common approach for studying reservoir parameters.Among these methods,deep learning methods are highly effective.From the perspective of multi-task learning,this paper uses six types of logging data—acoustic logging(AC),gamma ray(GR),compensated neutron porosity(CNL),density(DEN),deep and shallow lateral resistivity(LLD)and shallow lateral resistivity(LLS)—that are inputs and three reservoir parameters that are outputs to build a porosity saturation permeability network(PSP-Net)that can predict porosity,saturation,and permeability values simultaneously.These logging data are obtained from 108 training wells in a medium₋low permeability oilfield block in the western district of China.PSP-Net method adopts a serial structure to realize transfer learning of reservoir-parameter characteristics.Compared with other existing methods at the stage of academic exploration to simulating industrial applications,the proposed method overcomes the disadvantages inherent in single-task learning reservoir-parameter prediction models,including easily overfitting and heavy model-training workload.Additionally,the proposed method demonstrates good anti-overfitting and generalization capabilities,integrating professional knowledge and experience.In 37 test wells,compared with the existing method,the proposed method exhibited an average error reduction of 10.44%,27.79%,and 28.83%from porosity,saturation,permeability calculation.The prediction and actual permeabilities are within one order of magnitude.The training on PSP-Net are simpler and more convenient than other single-task learning methods discussed in this paper.Furthermore,the findings of this paper can help in the re-examination of old oilfield wells and the completion of logging data.展开更多
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.
基金supported by the National Natural Science Foundation of China(Grant Nos.42130719 and 42177173)the Doctoral Direct Train Project of Chongqing Natural Science Foundation(Grant No.CSTB2023NSCQ-BSX0029).
文摘Underground engineering projects such as deep tunnel excavation often encounter rockburst disasters accompanied by numerous microseismic events.Rapid interpretation of microseismic signals is crucial for the timely identification of rockbursts.However,conventional processing encompasses multi-step workflows,including classification,denoising,picking,locating,and computational analysis,coupled with manual intervention,which collectively compromise the reliability of early warnings.To address these challenges,this study innovatively proposes the“microseismic stethoscope"-a multi-task machine learning and deep learning model designed for the automated processing of massive microseismic signals.This model efficiently extracts three key parameters that are necessary for recognizing rockburst disasters:rupture location,microseismic energy,and moment magnitude.Specifically,the model extracts raw waveform features from three dedicated sub-networks:a classifier for source zone classification,and two regressors for microseismic energy and moment magnitude estimation.This model demonstrates superior efficiency compared to traditional processing and semi-automated processing,reducing per-event processing time from 0.71 s to 0.49 s to merely 0.036 s.It concurrently achieves 98%accuracy in source zone classification,with microseismic energy and moment magnitude estimation errors of 0.13 and 0.05,respectively.This model has been well applied and validated in the Daxiagu Tunnel case in Sichuan,China.The application results indicate that the model is as accurate as traditional methods in determining source parameters,and thus can be used to identify potential geomechanical processes of rockburst disasters.By enhancing the signal processing reliability of microseismic events,the proposed model in this study presents a significant advancement in the identification of rockburst disasters.
基金supported by the National Key R&D Program(No.2021YFB3501002)supported by the Ministry of Science and Technology of China,National Natural Science Foundation of China(No.51825101,52127801).
文摘Surface properties of crystals are critical in many fields,including electrochemistry and photoelectronics,the efficient prediction of which can expedite the design and optimization of catalysts,batteries,alloys etc.However,we are still far from realizing this vision due to the rarity of surface property-related databases,especially for multicomponent compounds,due to the large sample spaces and limited computing resources.In this work,we present a surface emphasized multi-task crystal graph convolutional neural network(SEM-CGCNN)to predict multiple surface properties simultaneously from crystal structures.The model is evaluated on a dataset of 3526 surface energies and work functions of binary magnesium intermetallics obtained through first-principles calculations,and obvious improvements are observed both in efficiency and accuracy over the original CGCNN model.By transferring the pre-trained model to the datasets of pure metals and other intermetallics,the fine-tuned SEM-CGCNN outperforms learning from scratch and can be further applied to other surface properties and materials systems.This study could be a paradigm for the end-to-end mapping of atomic structures to anisotropic surface properties of crystals,which provides an efficient framework to understand and screen materials with desired surface characteristics.
文摘Reconfigurable intelligent surface(RIS)have been cast as a promising alternative to alleviate blockage vulnerability and enhance coverage capability for terahertz(THz)communications.Owing to large-scale array elements at transceivers and RIS,the codebook based beamforming can be utilized in a computationally efficient manner.However,the codeword selection for analog beamforming is an intractable combinatorial optimization(CO)problem.To this end,by taking the CO problem as a classification problem,a multi-task learning based analog beam selection(MTL-ABS)framework is developed to implement cooperative beam selection concurrently at transceivers and RIS.In addition,residual network and self-attention mechanism are used to combat the network degradation and mine intrinsic THz channel features.Finally,the network convergence is analyzed from a blockwise perspective,and numerical results demonstrate that the MTL-ABS framework greatly decreases the beam selection overhead and achieves near optimal sum-rate compared with heuristic search based counterparts.
基金supported in part by the National Natural Science Foundation of China under Grant 62371181in part by the Changzhou Science and Technology International Cooperation Program under Grant CZ20230029The Institute of Information&Communications Technology Planning&Evaluation(IITP)grant funded by the Korea government(MSIT)(No.RS-2024-00396797,Development of core technology for intelligent O-RAN security platform).
文摘Satellite communication technology has emerged as a key solution to address the challenges of data transmission in remote areas.By overcoming the limitations of traditional terrestrial communication networks,it enables long-distance data transmission anytime and anywhere,ensuring the timely and accurate delivery of water level data,which is particularly crucial for fishway water level monitoring.To enhance the effectiveness of fishway water level monitoring,this study proposes a multi-task learning model,AS-SOMTF,designed for real-time and comprehensive prediction.The model integrates auxiliary sequences with primary input sequences to capture complex relationships and dependencies,thereby improving representational capacity.In addition,a novel timeseries embedding algorithm,AS-SOM,is introduced,which combines generative inference and pooling operations to optimize prediction efficiency for long sequences.This innovation not only ensures the timely transmission of water level data but also enhances the accuracy of real-time monitoring.Compared with traditional models such as Transformer and Long Short-Term Memory(LSTM)networks,the proposed model achieves improvements of 3.8%and 1.4%in prediction accuracy,respectively.These advancements provide more precise technical support for water level forecasting and resource management in the Diqing Tibetan Autonomous Prefecture of the Lancang River,contributing to ecosystem protection and improved operational safety.
文摘Deep neural networks(DNNs)are effective in solving both forward and inverse problems for nonlinear partial differential equations(PDEs).However,conventional DNNs are not effective in handling problems such as delay differential equations(DDEs)and delay integrodifferential equations(DIDEs)with constant delays,primarily due to their low regularity at delayinduced breaking points.In this paper,a DNN method that combines multi-task learning(MTL)which is proposed to solve both the forward and inverse problems of DIDEs.The core idea of this approach is to divide the original equation into multiple tasks based on the delay,using auxiliary outputs to represent the integral terms,followed by the use of MTL to seamlessly incorporate the properties at the breaking points into the loss function.Furthermore,given the increased training dificulty associated with multiple tasks and outputs,we employ a sequential training scheme to reduce training complexity and provide reference solutions for subsequent tasks.This approach significantly enhances the approximation accuracy of solving DIDEs with DNNs,as demonstrated by comparisons with traditional DNN methods.We validate the effectiveness of this method through several numerical experiments,test various parameter sharing structures in MTL and compare the testing results of these structures.Finally,this method is implemented to solve the inverse problem of nonlinear DIDE and the results show that the unknown parameters of DIDE can be discovered with sparse or noisy data.
基金The National Natural Science Foundation of China(62136008,62293541)The Beijing Natural Science Foundation(4232056)The Beijing Nova Program(20240484514).
文摘Cooperative multi-agent reinforcement learning(MARL)is a key technology for enabling cooperation in complex multi-agent systems.It has achieved remarkable progress in areas such as gaming,autonomous driving,and multi-robot control.Empowering cooperative MARL with multi-task decision-making capabilities is expected to further broaden its application scope.In multi-task scenarios,cooperative MARL algorithms need to address 3 types of multi-task problems:reward-related multi-task,arising from different reward functions;multi-domain multi-task,caused by differences in state and action spaces,state transition functions;and scalability-related multi-task,resulting from the dynamic variation in the number of agents.Most existing studies focus on scalability-related multitask problems.However,with the increasing integration between large language models(LLMs)and multi-agent systems,a growing number of LLM-based multi-agent systems have emerged,enabling more complex multi-task cooperation.This paper provides a comprehensive review of the latest advances in this field.By combining multi-task reinforcement learning with cooperative MARL,we categorize and analyze the 3 major types of multi-task problems under multi-agent settings,offering more fine-grained classifications and summarizing key insights for each.In addition,we summarize commonly used benchmarks and discuss future directions of research in this area,which hold promise for further enhancing the multi-task cooperation capabilities of multi-agent systems and expanding their practical applications in the real world.
基金supported by the research on key technologies for monitoring and identifying drug abuse of anesthetic drugs and psychotropic drugs,and intervention for addiction(No.2023YFC3304200)the program of a study on the diagnosis of addiction to synthetic cannabinoids and methods of assessing the risk of abuse(No.2022YFC3300905)+1 种基金the program of Ab initio design and generation of AI models for small molecule ligands based on target structures(No.2022PE0AC03)ZHIJIANG LAB.
文摘The accurate prediction of drug absorption,distribution,metabolism,excretion,and toxicity(ADMET)properties represents a crucial step in early drug development for reducing failure risk.Current deep learning approaches face challenges with data sparsity and information loss due to single-molecule representation limitations and isolated predictive tasks.This research proposes molecular properties prediction with parallel-view and collaborative learning(MolP-PC),a multi-view fusion and multi-task deep learning framework that integrates 1D molecular fingerprints(MFs),2D molecular graphs,and 3D geometric representations,incorporating an attention-gated fusion mechanism and multi-task adaptive learning strategy for precise ADMET property predictions.Experimental results demonstrate that MolP-PC achieves optimal performance in 27 of 54 tasks,with its multi-task learning(MTL)mechanism significantly enhancing predictive performance on small-scale datasets and surpassing single-task models in 41 of 54 tasks.Additional ablation studies and interpretability analyses confirm the significance of multi-view fusion in capturing multi-dimensional molecular information and enhancing model generalization.A case study examining the anticancer compound Oroxylin A demonstrates MolP-PC’s effective generalization in predicting key pharmacokinetic parameters such as half-life(T0.5)and clearance(CL),indicating its practical utility in drug modeling.However,the model exhibits a tendency to underestimate volume of distribution(VD),indicating potential for improvement in analyzing compounds with high tissue distribution.This study presents an efficient and interpretable approach for ADMET property prediction,establishing a novel framework for molecular optimization and risk assessment in drug development.
基金funded by the Excellent Talent Training Funding Project in Dongcheng District,Beijing,with project number 2024-dchrcpyzz-9.
文摘Predicting the health status of stroke patients at different stages of the disease is a critical clinical task.The onset and development of stroke are affected by an array of factors,encompassing genetic predisposition,environmental exposure,unhealthy lifestyle habits,and existing medical conditions.Although existing machine learning-based methods for predicting stroke patients’health status have made significant progress,limitations remain in terms of prediction accuracy,model explainability,and system optimization.This paper proposes a multi-task learning approach based on Explainable Artificial Intelligence(XAI)for predicting the health status of stroke patients.First,we design a comprehensive multi-task learning framework that utilizes the task correlation of predicting various health status indicators in patients,enabling the parallel prediction of multiple health indicators.Second,we develop a multi-task Area Under Curve(AUC)optimization algorithm based on adaptive low-rank representation,which removes irrelevant information from the model structure to enhance the performance of multi-task AUC optimization.Additionally,the model’s explainability is analyzed through the stability analysis of SHAP values.Experimental results demonstrate that our approach outperforms comparison algorithms in key prognostic metrics F1 score and Efficiency.
基金supported by the National Natural Science Foundation of China(Grant Nos.42030708,42375138,42030608,42105128,42075079)the Opening Foundation of Key Laboratory of Atmospheric Sounding,China Meteorological Administration(CMA),and the CMA Research Center on Meteorological Observation Engineering Technology(Grant No.U2021Z03),and the Opening Foundation of the Key Laboratory of Atmospheric Chemistry,CMA(Grant No.2022B02)。
文摘Aerosol optical depth(AOD)and fine particulate matter with a diameter of less than or equal to 2.5μm(PM_(2.5))play crucial roles in air quality,human health,and climate change.However,the complex correlation of AOD–PM_(2.5)and the limitations of existing algorithms pose a significant challenge in realizing the accurate joint retrieval of these two parameters at the same location.On this point,a multi-task learning(MTL)model,which enables the joint retrieval of PM_(2.5)concentration and AOD,is proposed and applied on the top-of-the-atmosphere reflectance data gathered by the Fengyun-4A Advanced Geosynchronous Radiation Imager(FY-4A AGRI),and compared to that of two single-task learning models—namely,Random Forest(RF)and Deep Neural Network(DNN).Specifically,MTL achieves a coefficient of determination(R^(2))of 0.88 and a root-mean-square error(RMSE)of 0.10 in AOD retrieval.In comparison to RF,the R^(2)increases by 0.04,the RMSE decreases by 0.02,and the percentage of retrieval results falling within the expected error range(Within-EE)rises by 5.55%.The R^(2)and RMSE of PM_(2.5)retrieval by MTL are 0.84 and 13.76μg m~(-3)respectively.Compared with RF,the R^(2)increases by 0.06,the RMSE decreases by 4.55μg m~(-3),and the Within-EE increases by 7.28%.Additionally,compared to DNN,MTL shows an increase of 0.01 in R^(2)and a decrease of 0.02 in RMSE in AOD retrieval,with a corresponding increase of 2.89%in Within-EE.For PM_(2.5)retrieval,MTL exhibits an increase of 0.05 in R^(2),a decrease of 1.76μg m~(-3)in RMSE,and an increase of 6.83%in Within-EE.The evaluation suggests that MTL is able to provide simultaneously improved AOD and PM_(2.5)retrievals,demonstrating a significant advantage in efficiently capturing the spatial distribution of PM_(2.5)concentration and AOD.
基金the National Key Research and Development Plan of China[Grant No.2023YFB3002400]the Shanghai 2021 Natural Science Foundation[Grant Nos.21ZR1420400 and 21ZR1419800]+1 种基金the Shanghai 2023 Natural Science Foundation[Grant No.23ZR1463000]the Shandong Provincial Meteorological Bureau Scientific Research Project[Grant No.2023SDBD05].
文摘Accurate short-term forecast of offshore wind fields is still challenging for numerical weather prediction models.Based on three years of 48-hour forecast data from the European Centre for Medium-Range Weather Forecasts Integrated Forecasting System global model(ECMWF-IFS)over 14 offshore weather stations along the coast of Shandong Province,this study introduces a multi-task learning(MTL)model(TabNet-MTL),which significantly improves the forecast bias of near-surface wind direction and speed simultaneously.TabNet-MTL adopts the feature engineering method,utilizes mean square error as the loss function,and employs the 5-fold cross validation method to ensure the generalization ability of the trained model.It demonstrates superior skills in wind field correction across different forecast lead times over all stations compared to its single-task version(TabNet-STL)and three other popular single-task learning models(Random Forest,LightGBM,and XGBoost).Results show that it significantly reduces root mean square error of the ECMWF-IFS wind speed forecast from 2.20 to 1.25 m s−1,and increases the forecast accuracy of wind direction from 50%to 65%.As an explainable deep learning model,the weather stations and long-term temporal statistics of near-surface wind speed are identified as the most influential variables for TabNet-MTL in constructing its feature engineering.
基金supported by the National Key Research and Development Program of China (2024YFD2001100,2024YFE0214300)the National Natural Science Foundation of China (62162008)+3 种基金Guizhou Provincial Science and Technology Projects ([2024]002, CXTD[2023]027)Guizhou Province Youth Science and Technology Talent Project ([2024]317)Guiyang Guian Science and Technology Talent Training Project ([2024]2-15)the Guizhou Province Graduate Education Innovation Program Project (2024YJSKYJJ096)
文摘Staple crops are the cornerstone of the food supply but are frequently threatened by plant diseases.Effective disease management,including disease identification and severity assessment,helps to better address these challenges.Currently,methods for disease severity assessment typically rely on calculating the area proportion of disease segmentation regions or using classification networks for severity assessment.However,these methods require large amounts of labeled data and fail to quantify lesion proportions when using classification networks,leading to inaccurate evaluations.To address these issues,we propose an automated framework for disease severity assessment that combines multi-task learning and knowledge-driven large-model segmentation techniques.This framework includes an image information processor,a lesion and leaf segmentation module,and a disease severity assessment module.First,the image information processor utilizes a multi-task learning strategy to analyze input images comprehensively,ensuring a deep understanding of disease characteristics.Second,the lesion and leaf segmentation module employ prompt-driven large-model technology to accurately segment diseased areas and entire leaves,providing detailed visual analysis.Finally,the disease severity assessment module objectively evaluates the severity of the disease based on professional grading standards by calculating lesion area proportions.Additionally,we have developed a comprehensive database of diseased leaf images from major crops,including several task-specific datasets.Experimental results demonstrate that our framework can accurately identify and assess the types and severity of crop diseases,even without extensive labeled data.Codes and data are available at http://dkp-ads.samlab.cn/.
基金supported by the Key Research and Development Program of Heilongjiang Province(No.2022ZX01A35).
文摘As the group-buying model shows significant progress in attracting new users,enhancing user engagement,and increasing platform profitability,providing personalized recommendations for group-buying users has emerged as a new challenge in the field of recommendation systems.This paper introduces a group-buying recommendation model based on multi-head attention mechanisms and multi-task learning,termed the Multi-head Attention Mechanisms and Multi-task Learning Group-Buying Recommendation(MAMGBR)model,specifically designed to optimize group-buying recommendations on e-commerce platforms.The core dataset of this study comes from the Chinese maternal and infant e-commerce platform“Beibei,”encompassing approximately 430,000 successful groupbuying actions and over 120,000 users.Themodel focuses on twomain tasks:recommending items for group organizers(Task Ⅰ)and recommending participants for a given group-buying event(Task Ⅱ).In model evaluation,MAMGBR achieves an MRR@10 of 0.7696 for Task I,marking a 20.23%improvement over baseline models.Furthermore,in Task II,where complex interaction patterns prevail,MAMGBR utilizes auxiliary loss functions to effectively model the multifaceted roles of users,items,and participants,leading to a 24.08%increase in MRR@100 under a 1:99 sample ratio.Experimental results show that compared to benchmark models,such as NGCF and EATNN,MAMGBR’s integration ofmulti-head attentionmechanisms,expert networks,and gating mechanisms enables more accurate modeling of user preferences and social associations within group-buying scenarios,significantly enhancing recommendation accuracy and platform group-buying success rates.
基金supported by the National Natural Science Foundation of China(No.52474435)China Baowu Low Carbon Metallurgy Innovation Foundation(BWLCF202307).
文摘Accurate forecasting of blast furnace gas(BFG)production is an essential prerequisite for reasonable energy scheduling and management to reduce carbon emissions.Coupling forecasting between BFG generation and consumption dynamics was taken as the research object.A multi-task learning(MTL)method for BFG forecasting was proposed,which integrated a coupling correlation coefficient(CCC)and an inverted transformer structure.The CCC method could enhance key information extraction by establishing relationships between multiple prediction targets and relevant factors,while MTL effectively captured the inherent correlations between BFG generation and consumption.Finally,a real-world case study was conducted to compare the proposed model with four benchmark models.Results indicated significant reductions in average mean absolute percentage error by 33.37%,achieving 1.92%,with a computational time of 76 s.The sensitivity analysis of hyperparameters such as learning rate,batch size,and units of the long short-term memory layer highlights the importance of hyperparameter tuning.
文摘With focus now placed on the learner, more attention is given to his learning style, multiple intelligence and developing learning strategies to enable him to make sense of and use of the target language appropriately in varied contexts and with different uses of the language. To attain this, the teacher is tasked with designing, monitoring and processing language learning activities for students to carry out and in the process learn by doing and reflecting on the learning process they went through as they interacted socially with each other. This paper describes a task named"The Fishbowl Technique"and found to be effective in large ESL classes in the secondary level in the Philippines.
基金supported by the National Natural Science Foundation of China(Nos.1187050492,12005303,and 12175170).
文摘The global nuclear mass based on the macroscopic-microscopic model was studied by applying a newly designed multi-task learning artificial neural network(MTL-ANN). First, the reported nuclear binding energies of 2095 nuclei(Z ≥ 8, N ≥ 8) released in the latest Atomic Mass Evaluation AME2020 and the deviations between the fitting result of the liquid drop model(LDM)and data from AME2020 for each nucleus were obtained.To compensate for the deviations and investigate the possible ignored physics in the LDM, the MTL-ANN method was introduced in the model. Compared to the single-task learning(STL) method, this new network has a powerful ability to simultaneously learn multi-nuclear properties,such as the binding energies and single neutron and proton separation energies. Moreover, it is highly effective in reducing the risk of overfitting and achieving better predictions. Consequently, good predictions can be obtained using this nuclear mass model for both the training and validation datasets and for the testing dataset. In detail, the global root mean square(RMS) of the binding energy is effectively reduced from approximately 2.4 MeV of LDM to the current 0.2 MeV, and the RMS of Sn, Spcan also reach approximately 0.2 MeV. Moreover, compared to STL, for the training and validation sets, 3-9% improvement can be achieved with the binding energy, and 20-30% improvement for S_(n), S_(p);for the testing sets, the reduction in deviations can even reach 30-40%, which significantly illustrates the advantage of the current MTL.
基金National Key R&D Program of China under Grant No.2019YFC1511005the National Natural Science Foundation of China under Grant Nos.51921006,52192661 and 52008138+2 种基金the China Postdoctoral Science Foundation under Grant Nos.BX20190102 and 2019M661286the Heilongjiang Natural Science Foundation under Grant No.LH2022E070the Heilongjiang Province Postdoctoral Science Foundation under Grant Nos.LBH-TZ2016 and LBH-Z19064。
文摘Recent studies for computer vision and deep learning-based,post-earthquake inspections on RC structures mainly perform well for specific tasks,while the trained models must be fine-tuned and re-trained when facing new tasks and datasets,which is inevitably time-consuming.This study proposes a multi-task learning approach that simultaneously accomplishes the semantic segmentation of seven-type structural components,three-type seismic damage,and four-type deterioration states.The proposed method contains a CNN-based encoder-decoder backbone subnetwork with skip-connection modules and a multi-head,task-specific recognition subnetwork.The backbone subnetwork is designed to extract multi-level features of post-earthquake RC structures.The multi-head,task-specific recognition subnetwork consists of three individual self-attention pipelines,each of which utilizes extracted multi-level features from the backbone network as a mutual guidance for the individual segmentation task.A synthetical loss function is designed with real-time adaptive coefficients to balance multi-task losses and focus on the most unstably fluctuating one.Ablation experiments and comparative studies are further conducted to demonstrate their effectiveness and necessity.The results show that the proposed method can simultaneously recognize different structural components,seismic damage,and deterioration states,and that the overall performance of the three-task learning models gains general improvement when compared to all single-task and dual-task models.
基金supported by the Science and Technology Grant No.520120210003,Jibei Electric Power Company of the State Grid Corporation of China。
文摘Convective storms and lightning are among the most important weather phenomena that are challenging to forecast.In this study,a novel multi-task learning(MTL)encoder-decoder U-net neural network was developed to forecast convective storms and lightning with lead times for up to 90 min,using GOES-16 geostationary satellite infrared brightness temperatures(IRBTs),lightning flashes from Geostationary Lightning Mapper(GLM),and vertically integrated liquid(VIL)from Next Generation Weather Radar(NEXRAD).To cope with the heavily skewed distribution of lightning data,a spatiotemporal exponent-weighted loss function and log-transformed lightning normalization approach were developed.The effects of MTL,single-task learning(STL),and IRBTs as auxiliary input features on convection and lightning nowcasting were investigated.The results showed that normalizing the heavily skew-distributed lightning data along with a log-transformation dramatically outperforms the min-max normalization method for nowcasting an intense lightning event.The MTL model significantly outperformed the STL model for both lightning nowcasting and VIL nowcasting,particularly for intense lightning events.The MTL also helped delay the lightning forecast performance decay with the lead times.Furthermore,incorporating satellite IRBTs as auxiliary input features substantially improved lightning nowcasting,but produced little difference in VIL forecasting.Finally,the MTL model performed better for forecasting both lightning and the VIL of organized convective storms than for isolated cells.
文摘There is a growing amount of data uploaded to the internet every day and it is important to understand the volume of those data to find a better scheme to process them.However,the volume of internet data is beyond the processing capabilities of the current internet infrastructure.Therefore,engineering works using technology to organize and analyze information and extract useful information are interesting in both industry and academia.The goal of this paper is to explore the entity relationship based on deep learning,introduce semantic knowledge by using the prepared language model,develop an advanced entity relationship information extraction method by combining Robustly Optimized BERT Approach(RoBERTa)and multi-task learning,and combine the intelligent characters in the field of linguistic,called Robustly Optimized BERT Approach+Multi-Task Learning(RoBERTa+MTL).To improve the effectiveness of model interaction,multi-task teaching is used to implement the observation information of auxiliary tasks.Experimental results show that our method has achieved an accuracy of 88.95 entity relationship extraction,and a further it has achieved 86.35%of accuracy after being combined with multi-task learning.
文摘Buiding data-driven models using machine learning methods has gradually become a common approach for studying reservoir parameters.Among these methods,deep learning methods are highly effective.From the perspective of multi-task learning,this paper uses six types of logging data—acoustic logging(AC),gamma ray(GR),compensated neutron porosity(CNL),density(DEN),deep and shallow lateral resistivity(LLD)and shallow lateral resistivity(LLS)—that are inputs and three reservoir parameters that are outputs to build a porosity saturation permeability network(PSP-Net)that can predict porosity,saturation,and permeability values simultaneously.These logging data are obtained from 108 training wells in a medium₋low permeability oilfield block in the western district of China.PSP-Net method adopts a serial structure to realize transfer learning of reservoir-parameter characteristics.Compared with other existing methods at the stage of academic exploration to simulating industrial applications,the proposed method overcomes the disadvantages inherent in single-task learning reservoir-parameter prediction models,including easily overfitting and heavy model-training workload.Additionally,the proposed method demonstrates good anti-overfitting and generalization capabilities,integrating professional knowledge and experience.In 37 test wells,compared with the existing method,the proposed method exhibited an average error reduction of 10.44%,27.79%,and 28.83%from porosity,saturation,permeability calculation.The prediction and actual permeabilities are within one order of magnitude.The training on PSP-Net are simpler and more convenient than other single-task learning methods discussed in this paper.Furthermore,the findings of this paper can help in the re-examination of old oilfield wells and the completion of logging data.
