"Towards Efficient Architectural Support for AI-based IoT Applications"

Mingcong Song, Ph.D. candidate, ECE, University of Florida

April 4th (Wednesday), 11:00am
Harold Frank Hall (HFH), Rm. 4164 (ECE Conf. Rm.)

In recent years, the artificial intelligence (AI) techniques, represented by deep neural networks (DNN), have demonstrated transformative impacts to modern Internet-of-Things (IoT) applications such as smart cities and smart transportation. With the increasing computing power and energy efficiency of mobile devices, there is a growing interest in performing AI-based IoT applications on mobile platforms. As a result, we believe the next-generation AI-based applications are pervasive across all platforms, ranging from central cloud data center to edge-side wearable and mobile devices.

However, we observe several architectural gaps that challenge the pervasive AI. First, the diversity of computing hardware resources and different end-user requirements present challenges to AI-based applications deployment on various IoT platforms, which results in inferior user satisfaction. Second, the traditional statically trained DNN model could not efficiently handle the dynamic data in the real IoT environments, which leads to low inference accuracy. Lastly, the training of DNN models still involves extensive human efforts to collect and label the large-scale dataset, which becomes impractical in IoT big data era where raw IoT data is largely un-labeled and un-categorized.

In this talk, I will introduce my research which enables pervasive AI-based IoT applications to become high-efficient, user-satisfactory, and intelligent. I will first introduce Pervasive AI, a user satisfaction-aware deep learning inference framework, to provide the best user satisfaction when migrating AI-based applications from Cloud to all kinds of platforms. Next, I will describe In-situ AI, a novel computing paradigm tailored to AI-based IoT applications. Finally, to achieve real intelligent (support autonomous learning) in IoT nodes, I will introduce an unsupervised GAN-based deep learning accelerator.

About Mingcong Song:

Mingcong Song is a Ph.D. candidate in the Department of Electrical and Computer Engineering at the University of Florida. His research interests include architectural support for emerging AI applications, AI-enabled IoT system design and heterogeneous computing for big data applications. His work has been published in top-tier conferences including ISCA, HPCA, ASPLOS, PACT, ICS, etc. His research has won the best paper nomination at HPCA 2017. He received a BS degree from Huazhong University of Science and Technology in 2010 and an MS degree from University of Chinese Academy Sciences in 2013.

Hosted by: UCSB CE Program