Home > Events > Oral Candidacy - Qiuwen Lou

Oral Candidacy - Qiuwen Lou

Start: 12/7/2017 at 12:30PM
End: 12/7/2017 at 3:30PM
Location: 258 Fitzpatrick Hall
Add to calendar:
iCal vCal

Qiuwen Lou

Oral Candidacy
December 7, 2017        12:30 pm        258 Fitzpatrick

Adviser:  Dr. Sharon Hu
Dr. Azad Naeemi        Dr. Walter Scheirer        Dr. Yiyu Shi


 Non von Neumann Architecture based Image Processing Primitives


 With the success of many computer vision algorithms proposed nowadays, researchers are calling for efficient computation paradigm to accelerate these computations. Some non von Neumann architectures demonstrate promising candidates to transcend the performance/energy limitation. Previous research has shown this kind of architecture is a powerful processor that can significantly improve the performance of spatio-temporal applications when compared to the more traditional approach. In this proposal, we focus on the cellular neural network (CeNN) and its inspired architecture for computer vision application acceleration. First, we show our CeNN design using both conventional CMOS and emerging technologies. Then, we demonstrate several application level study based on cellular neural network and its inspired architecture, including tracking and convolutional neural networks. We designed algorithms that are CeNN friendly, and uses CeNN based architecture for benchmarking. Our result indicate that such software/hardware co-design methodology can benefit various applications. We obtained 100x to 1000x energy-delay product (EDP) benefits compared with other approaches. We also proposed our future work. It mainly focuses on three folds. (1). Explore larger convolutional neural network with more layers and feature maps. 2). Explore novel training method to improve the accuracy for analog based deep neural network implementation. 3). Explore other type of deep neural networks, including ResNet and recurrent neural networks.