Home > Events > Oral Candidacy - Jinglan Liu

Oral Candidacy - Jinglan Liu

Start: 1/15/2019 at 3:00PM
End: 1/15/2019 at 5:30PM
Location: 315 Stinson Remick
Attendees: Faculty and students are welcome to attend the presentation portion of the defense.
Add to calendar:
iCal vCal

Jinglan Liu

Oral Candidacy Exam

January 15, 2019      3:00 pm      315 Stinson Remick

Adviser:  Dr. Yiyu Shi

Committee Members:

Dr. Sharon Hu      Dr. Micheal Niemier      Dr. Jinjun Xiong



Acceleration and Applications for Generative Adversarial Networks


In recent years, generative adversarial networks (GANs), which are spin-offs from conventional convolutional neural networks (CNNs), have attracted much attention in the fields of reinforcement learning, unsupervised learning and also semi-supervised learning. To solve the high computational and memory cost problem inherited from CNN, compression techniques such as using binary weights instead of floating-point numbers can be readily applied to discriminator networks in GANs; However, adopting it directly will fail generator networks. We solved this problem by adaptive partial binarization. 

In addition, I attempted to extend GANs into two promising applications for further speedup. I have successfully applied Multi-Cycle GANs to the CT (computed tomography) artifacts removal problem. It is always desired to find a approach that removes the artifacts in CT images to make them as clear as possible for doctors. Multi-Cycle GAN solved this task gradually. Upon similar training and testing time, Multi-Cycle GANs can achieve better CT scan image quality than vanilla CycleGANs, which is the state-of-the-art method in academia. 

I propose to apply GAN into helping noise maps simulation. The relentless efforts towards power reduction of integrated circuits have led to the prevalence of near-threshold computing paradigms. As a result, designers seek to contain noise violations, commonly known as voltage emergencies, through various techniques. All these techniques require accurate capture of voltage emergencies through noise sensors, which is based on a large amount of circuit simulations. Nevertheless, circuit simulation is extremely time-consuming, and it usually costs several weeks to produce enough simulated samples for the noise sensor placement decision. I will investigate an approach based on GANs, which greatly reduced the time needed to generate plenty samples for a good noise sensor placement solution.