Knowledge distillation(KD)is an emerging model compression technique for learning compact object detector models.Previous KD often focused solely on distilling from the logits layer or the feature intermediate layers,...Knowledge distillation(KD)is an emerging model compression technique for learning compact object detector models.Previous KD often focused solely on distilling from the logits layer or the feature intermediate layers,which may limit the comprehensive learning of the student network.Additionally,the imbalance between the foreground and background also affects the performance of the model.To address these issues,this paper employs feature-based distillation to enhance the detection performance of the bounding box localization part,and logit-based distillation to improve the detection performance of the category prediction part.Specifically,for the intermediate layer feature distillation,we introduce feature resampling to reduce the risk of the student model merely imitating the teacher model.At the same time,we incorporate a Spatial Attention Mechanism(SAM)to highlight the foreground features learned by the student model.In terms of output layer feature distillation,we divide the traditional distillation targets into target-class objects and non-target-class objects,aiming to improve overall distillation performance.Furthermore,we introduce a one-to-many matching distillation strategy based on Feature Alignment Module(FAM),which further enhances the studentmodel’s feature representation ability,making its feature distribution closer to that of the teacher model,and thus demonstrating superior localization and classification capabilities in object detection tasks.Experimental results demonstrate that our proposedmethodology outperforms conventional distillation techniques in terms of object detecting performance.展开更多
Imputation of missing data has long been an important topic and an essential application for intelligent transportation systems(ITS)in the real world.As a state-of-the-art generative model,the diffusion model has prov...Imputation of missing data has long been an important topic and an essential application for intelligent transportation systems(ITS)in the real world.As a state-of-the-art generative model,the diffusion model has proven highly successful in image generation,speech generation,time series modelling etc.and now opens a new avenue for traffic data imputation.In this paper,we propose a conditional diffusion model,called the implicit-explicit diffusion model,for traffic data imputation.This model exploits both the implicit and explicit feature of the data simultaneously.More specifically,we design two types of feature extraction modules,one to capture the implicit dependencies hidden in the raw data at multiple time scales and the other to obtain the long-term temporal dependencies of the time series.This approach not only inherits the advantages of the diffusion model for estimating missing data,but also takes into account the multiscale correlation inherent in traffic data.To illustrate the performance of the model,extensive experiments are conducted on three real-world time series datasets using different missing rates.The experimental results demonstrate that the model improves imputation accuracy and generalization capability.展开更多
隐喻是人类语言中经常出现的一种特殊现象,隐喻识别对于自然语言处理各项任务来说具有十分基础和重要的意义。针对中文领域的隐喻识别任务,该文提出了一种基于句法感知图卷积神经网络和ELECTRA的隐喻识别模型(S yntax-a ware G CN with ...隐喻是人类语言中经常出现的一种特殊现象,隐喻识别对于自然语言处理各项任务来说具有十分基础和重要的意义。针对中文领域的隐喻识别任务,该文提出了一种基于句法感知图卷积神经网络和ELECTRA的隐喻识别模型(S yntax-a ware G CN with E LECTRA,SaGE)。该模型从语言学出发,使用ELECTRA和Transformer编码器抽取句子的语义特征,将句子按照依存关系组织成一张图并使用图卷积神经网络抽取其句法特征,在此基础上对两类特征进行融合以进行隐喻识别。该模型在CCL 2018中文隐喻识别评测数据集上以85.22%的宏平均F 1值超越了此前的最佳成绩,验证了融合语义信息和句法信息对于隐喻识别任务具有重要作用。展开更多
基金funded by National Natural Science Foundation of China(61603245).
文摘Knowledge distillation(KD)is an emerging model compression technique for learning compact object detector models.Previous KD often focused solely on distilling from the logits layer or the feature intermediate layers,which may limit the comprehensive learning of the student network.Additionally,the imbalance between the foreground and background also affects the performance of the model.To address these issues,this paper employs feature-based distillation to enhance the detection performance of the bounding box localization part,and logit-based distillation to improve the detection performance of the category prediction part.Specifically,for the intermediate layer feature distillation,we introduce feature resampling to reduce the risk of the student model merely imitating the teacher model.At the same time,we incorporate a Spatial Attention Mechanism(SAM)to highlight the foreground features learned by the student model.In terms of output layer feature distillation,we divide the traditional distillation targets into target-class objects and non-target-class objects,aiming to improve overall distillation performance.Furthermore,we introduce a one-to-many matching distillation strategy based on Feature Alignment Module(FAM),which further enhances the studentmodel’s feature representation ability,making its feature distribution closer to that of the teacher model,and thus demonstrating superior localization and classification capabilities in object detection tasks.Experimental results demonstrate that our proposedmethodology outperforms conventional distillation techniques in terms of object detecting performance.
基金partially supported by the National Natural Science Foundation of China(62271485)the SDHS Science and Technology Project(HS2023B044)
文摘Imputation of missing data has long been an important topic and an essential application for intelligent transportation systems(ITS)in the real world.As a state-of-the-art generative model,the diffusion model has proven highly successful in image generation,speech generation,time series modelling etc.and now opens a new avenue for traffic data imputation.In this paper,we propose a conditional diffusion model,called the implicit-explicit diffusion model,for traffic data imputation.This model exploits both the implicit and explicit feature of the data simultaneously.More specifically,we design two types of feature extraction modules,one to capture the implicit dependencies hidden in the raw data at multiple time scales and the other to obtain the long-term temporal dependencies of the time series.This approach not only inherits the advantages of the diffusion model for estimating missing data,but also takes into account the multiscale correlation inherent in traffic data.To illustrate the performance of the model,extensive experiments are conducted on three real-world time series datasets using different missing rates.The experimental results demonstrate that the model improves imputation accuracy and generalization capability.
文摘隐喻是人类语言中经常出现的一种特殊现象,隐喻识别对于自然语言处理各项任务来说具有十分基础和重要的意义。针对中文领域的隐喻识别任务,该文提出了一种基于句法感知图卷积神经网络和ELECTRA的隐喻识别模型(S yntax-a ware G CN with E LECTRA,SaGE)。该模型从语言学出发,使用ELECTRA和Transformer编码器抽取句子的语义特征,将句子按照依存关系组织成一张图并使用图卷积神经网络抽取其句法特征,在此基础上对两类特征进行融合以进行隐喻识别。该模型在CCL 2018中文隐喻识别评测数据集上以85.22%的宏平均F 1值超越了此前的最佳成绩,验证了融合语义信息和句法信息对于隐喻识别任务具有重要作用。
