Song Tao

Mr. Song Tao

Shanghai Jiao Tong University

Research Focus & Metrics

Academic Career Length

8Years (2019-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
3 July 2019 in IEEE Transactions on Signal Processing
Verified
Journal Article
Closed access
YXYili Xia
Yili Xia
STSong Tao
Song Tao
ZLZhe Li
Zhe Li
Min XiangMin Xiang
Min Xiang
WPWenjiang Pei
Wenjiang Pei
DMDanilo Mandic
Danilo Mandic
Novel physical insights are provided into the mean square performance bounds of quaternion-valued widely linear (WL), semi-widely linear (SWL), and strictly linear (SL) estimators for the generality of quaternion-valued Gaussian data. This is achieved by first defining three kinds of complementary mean square errors (CMSEs) of these estimators and by further exploiting the corresponding degrees of H-improperness (second-order noncircularity). Next, the investigation of the bounds of the attainable CMSEs by these classes of estimators shows that only a joint consideration of the proposed CMSE analysis and the standard MSE analysis provides enough degrees of freedom for a detailed account of the MSE performance. The so established framework for the analysis of estimators for H-improper data is shown to be capable of measuring error power distribution for each data channel, an important finding which is not possible to obtain through the standard MSE analysis only. Simulations in the system identification setting support the analysis.

16 March 2020 in arXiv
Verified
Preprint
Open access
PDF file
STSong Tao
Song Tao
JWJia Wang
Jia Wang
In order to alleviate the notorious mode collapse phenomenon in generative adversarial networks (GANs), we propose a novel training method of GANs in which certain fake samples are considered as real ones during the training process. This strategy can reduce the gradient value that generator receives in the region where gradient exploding happens. We show the process of an unbalanced generation and a vicious circle issue resulted from gradient exploding in practical training, which explains the instability of GANs. We also theoretically prove that gradient exploding can be alleviated by penalizing the difference between discriminator outputs and fake-as-real consideration for very close real and fake samples. Accordingly, Fake-As-Real GAN (FARGAN) is proposed with a more stable training process and a more faithful generated distribution. Experiments on different datasets verify our theoretical analysis.