The lack of semantic parts, increasing the number of Web services in the Web, and syntactic-based search operation are the main problems of current Web service technologies, these factors make difficult for clients to...The lack of semantic parts, increasing the number of Web services in the Web, and syntactic-based search operation are the main problems of current Web service technologies, these factors make difficult for clients to find a required web service. This paper shows a matchmaking algorithm to discover Semantic Web Services that are satisfying client requirements. It depends on two factors that distinguish it from any conventional Web service discovery algorithm;the first one is using semantic matching technique to overcome shortcoming of keyword matching techniques, the second one is tying Quality of Service (QoS) metrics of Web Service (WS) with fuzzy words that are used in user’s request. At least fifty percent average gain in search relevancy is obtained when our matchmaking algorithm is applied to WSs that are actually matching the chosen fuzzy semantic theme.展开更多
Frequent counting is a very so often required operation in machine learning algorithms. A typical machine learning task, learning the structure of Bayesian network （BN） based on metric scoring, is introduced as an e...Frequent counting is a very so often required operation in machine learning algorithms. A typical machine learning task, learning the structure of Bayesian network （BN） based on metric scoring, is introduced as an example that heavily relies on frequent counting. A fast calculation method for frequent counting enhanced with two cache layers is then presented for learning BN. The main contribution of our approach is to eliminate comparison operations for frequent counting by introducing a multi-radix number system calculation. Both mathematical analysis and empirical comparison between our method and state-of-the-art solution are conducted. The results show that our method is dominantly superior to state-of-the-art solution in solving the problem of learning BN.展开更多
传统的基于深度学习的图像分类方法在大样本分类任务中具有较好的分类效果,但在小样本分类任务中却存在较大的挑战,为此,小样本图像分类获得了研究人员的广泛关注.基于度量的方法是解决小样本图像分类的一种简单有效方法,它利用可学习...传统的基于深度学习的图像分类方法在大样本分类任务中具有较好的分类效果,但在小样本分类任务中却存在较大的挑战,为此,小样本图像分类获得了研究人员的广泛关注.基于度量的方法是解决小样本图像分类的一种简单有效方法,它利用可学习的映射函数将分类任务中的所有样本映射到一个特征空间中,然后基于某种度量标准对查询特征进行分类.由于分类任务中不同类的两个图像有可能包含较多的相似性区域,导致特征空间中某些查询特征与异类的类原型特征的距离较近,较难学习到大的分类边界.为了解决上述问题,本文提出了注意力全关系网络(Total Relation Network with Attention,TRNA),该网络通过计算特征对的全关系和特征对的注意力来实现大边界的特征空间.具体地,在计算出所有的查询特征和类原型后,提出的网络利用特征对全关系拼接操作将特征空间中的任意两个特征在通道方向上进行拼接得到特征对矩阵,然后利用特征对注意力机制将特征对矩阵中不同类间难区分的特征对挑选出来并给予大的权重,最后将特征对矩阵输入卷积网络和全连接网络得到一个相似得分矩阵.实验结果表明本文的方法与关系网络相比,在数据集mini-ImageNet、Stanford-Dogs、Stanford-Cars、CUB-200-2011的1-shot和5-shot分类任务中分别有2.67%和1.71%、8.31%和3.92%、14.99%和8.00%、4.41%和4.42%的性能提升.展开更多
文摘The lack of semantic parts, increasing the number of Web services in the Web, and syntactic-based search operation are the main problems of current Web service technologies, these factors make difficult for clients to find a required web service. This paper shows a matchmaking algorithm to discover Semantic Web Services that are satisfying client requirements. It depends on two factors that distinguish it from any conventional Web service discovery algorithm;the first one is using semantic matching technique to overcome shortcoming of keyword matching techniques, the second one is tying Quality of Service (QoS) metrics of Web Service (WS) with fuzzy words that are used in user’s request. At least fifty percent average gain in search relevancy is obtained when our matchmaking algorithm is applied to WSs that are actually matching the chosen fuzzy semantic theme.
基金supported by National Natural Science Foundation of China (No.60970055)
文摘Frequent counting is a very so often required operation in machine learning algorithms. A typical machine learning task, learning the structure of Bayesian network （BN） based on metric scoring, is introduced as an example that heavily relies on frequent counting. A fast calculation method for frequent counting enhanced with two cache layers is then presented for learning BN. The main contribution of our approach is to eliminate comparison operations for frequent counting by introducing a multi-radix number system calculation. Both mathematical analysis and empirical comparison between our method and state-of-the-art solution are conducted. The results show that our method is dominantly superior to state-of-the-art solution in solving the problem of learning BN.
文摘传统的基于深度学习的图像分类方法在大样本分类任务中具有较好的分类效果,但在小样本分类任务中却存在较大的挑战,为此,小样本图像分类获得了研究人员的广泛关注.基于度量的方法是解决小样本图像分类的一种简单有效方法,它利用可学习的映射函数将分类任务中的所有样本映射到一个特征空间中,然后基于某种度量标准对查询特征进行分类.由于分类任务中不同类的两个图像有可能包含较多的相似性区域,导致特征空间中某些查询特征与异类的类原型特征的距离较近,较难学习到大的分类边界.为了解决上述问题,本文提出了注意力全关系网络(Total Relation Network with Attention,TRNA),该网络通过计算特征对的全关系和特征对的注意力来实现大边界的特征空间.具体地,在计算出所有的查询特征和类原型后,提出的网络利用特征对全关系拼接操作将特征空间中的任意两个特征在通道方向上进行拼接得到特征对矩阵,然后利用特征对注意力机制将特征对矩阵中不同类间难区分的特征对挑选出来并给予大的权重,最后将特征对矩阵输入卷积网络和全连接网络得到一个相似得分矩阵.实验结果表明本文的方法与关系网络相比,在数据集mini-ImageNet、Stanford-Dogs、Stanford-Cars、CUB-200-2011的1-shot和5-shot分类任务中分别有2.67%和1.71%、8.31%和3.92%、14.99%和8.00%、4.41%和4.42%的性能提升.
