卷积神经网络已在多个领域取得了优异的性能表现,然而由于其不透明的内部状态,其可解释性依然面临很大的挑战.其中一个原因是卷积神经网络以像素级特征为输入,逐层地抽取高级别特征,然而这些高层特征依然十分抽象,人类不能直观理解.为...卷积神经网络已在多个领域取得了优异的性能表现,然而由于其不透明的内部状态,其可解释性依然面临很大的挑战.其中一个原因是卷积神经网络以像素级特征为输入,逐层地抽取高级别特征,然而这些高层特征依然十分抽象,人类不能直观理解.为了解决这一问题,我们需要表征出网络中隐藏的人类可理解的语义概念.本文通过预先定义语义概念数据集(例如红色、条纹、斑点、狗),得到这些语义在网络某一层的特征图,将这些特征图作为数据,训练一个张量分类器.我们将与分界面正交的张量称为语义激活张量(Semantic Activation Tensors,SATs),每个SAT都指向对应的语义概念.相对于向量分类器,张量分类器可以保留张量数据的原始结构.在卷积网络中,每个特征图中都包含了位置信息和通道信息,如果将其简单地展开成向量形式,这会破坏其结构信息,导致最终分类精度的降低.本文使用SAT与网络梯度的内积来量化语义对分类结果的重要程度,此方法称为TSAT(Testing with SATs).例如,条纹对斑马的预测结果有多大影响.本文以图像分类网络作为解释对象,数据集选取ImageNet,在ResNet50和Inceptionv3两种网络架构上进行实验验证.最终实验结果表明,本文所采用的张量分类方法相较于传统的向量分类方法,在数据维度较大或数据不易区分的情况下,分类精度有显著的提高,且分类的稳定性也更加优秀.这从而保证了本文所推导出的语义激活张量更加准确,进一步确保了后续语义概念重要性量化的准确性.展开更多
Many offshore marine structures are built on the seabed that are slightly or considerably sloping.To study the sloping seabed transient response during marine earthquakes,an analytical solution induced by a P-wave lin...Many offshore marine structures are built on the seabed that are slightly or considerably sloping.To study the sloping seabed transient response during marine earthquakes,an analytical solution induced by a P-wave line source embedded in the solid is presented.During the derivation,the wave fields in the fluid layer and the semi-infinite solid are firstly constructed by using the generalized ray method and the fluid-solid interface reflection and transmission coefficients.Then,the analytical solution in the transformed domain is obtained by superposing these wave fields,and the analytical solution in the time domain by applying the analytical inverse Laplace transform method.The the head wave generation conditions and arrival times at the fluid-solid interface are derived through this solution.Through the use of numerical examples,the analytical solution is proved right and the impacts of the sloping angle on the hydrodynamic pressure in the sea,the seismic wave propagation in the seabed,the head wave,and the Scholte wave at the seawater-seabed interface are also addressed.展开更多
文摘卷积神经网络已在多个领域取得了优异的性能表现,然而由于其不透明的内部状态,其可解释性依然面临很大的挑战.其中一个原因是卷积神经网络以像素级特征为输入,逐层地抽取高级别特征,然而这些高层特征依然十分抽象,人类不能直观理解.为了解决这一问题,我们需要表征出网络中隐藏的人类可理解的语义概念.本文通过预先定义语义概念数据集(例如红色、条纹、斑点、狗),得到这些语义在网络某一层的特征图,将这些特征图作为数据,训练一个张量分类器.我们将与分界面正交的张量称为语义激活张量(Semantic Activation Tensors,SATs),每个SAT都指向对应的语义概念.相对于向量分类器,张量分类器可以保留张量数据的原始结构.在卷积网络中,每个特征图中都包含了位置信息和通道信息,如果将其简单地展开成向量形式,这会破坏其结构信息,导致最终分类精度的降低.本文使用SAT与网络梯度的内积来量化语义对分类结果的重要程度,此方法称为TSAT(Testing with SATs).例如,条纹对斑马的预测结果有多大影响.本文以图像分类网络作为解释对象,数据集选取ImageNet,在ResNet50和Inceptionv3两种网络架构上进行实验验证.最终实验结果表明,本文所采用的张量分类方法相较于传统的向量分类方法,在数据维度较大或数据不易区分的情况下,分类精度有显著的提高,且分类的稳定性也更加优秀.这从而保证了本文所推导出的语义激活张量更加准确,进一步确保了后续语义概念重要性量化的准确性.
基金financially supported by the National Key R&D Program of China (Grant No.2021YFC3100700)the National Natural Science Foundation of China (Grant Nos.U2039209 and 41874067)the Natural Science Foundation of Heilongjiang Province,China (Grant No.YQ2021D010)。
文摘Many offshore marine structures are built on the seabed that are slightly or considerably sloping.To study the sloping seabed transient response during marine earthquakes,an analytical solution induced by a P-wave line source embedded in the solid is presented.During the derivation,the wave fields in the fluid layer and the semi-infinite solid are firstly constructed by using the generalized ray method and the fluid-solid interface reflection and transmission coefficients.Then,the analytical solution in the transformed domain is obtained by superposing these wave fields,and the analytical solution in the time domain by applying the analytical inverse Laplace transform method.The the head wave generation conditions and arrival times at the fluid-solid interface are derived through this solution.Through the use of numerical examples,the analytical solution is proved right and the impacts of the sloping angle on the hydrodynamic pressure in the sea,the seismic wave propagation in the seabed,the head wave,and the Scholte wave at the seawater-seabed interface are also addressed.
