学术交流

新葡萄娱乐官网学术报告 第2020-06-12期

发布时间:2020-06-12动态浏览次数:17

Title

Sustainable AI

Lecture & Organization

Di Wang

Principal Research Lead at Microsoft Mixed Reality

Date & Time

1:00-2:00pm  Fri, June 12, 2020

Beijing time

Mode

Zoom ID:  574 459 9066

Password: 001234

Bio

Dr. Di Wang is currently a Principal Research lead at Microsoft Mixed Reality organization. He received his Ph.D. in Computer Science and Engineering from The Pennsylvania State University in 2014, M.S. in Computer Systems Engineering from Technical University of Denmark (DTU) in 2008 and B.E. in Computer Science and Technology from Zhejiang University in 2005. He joined Microsoft Research in 2014 and has worked on various research projects and products including AI based solutions, cloud infrastructure optimization and IoT. 

His research work spans the areas of AI/machine learning/deep learning, computer systems and architecture, energy-efficient systems design and sustainable computing, software hardware co-design, and VLSI design. Specifically, he has applied his expertise on these topics to the areas of AI systems, datacenters, IoT, storage systems, fault tolerant systems, and EDA tools. He has authored over 40 publications in top conferences and journals and has received 7 best paper awards/nominations. In addition, he has filed more than 10 patents at Microsoft. His work has also been featured in the CACM news and was chosen as IEEE sustainable computing register’s pick of the month. He is an active member in the research community serving as program/track Chair and/or program committee member of several top-notch conferences and frequent reviewers of several journals.

Abstract

The success of Deep Neural Networks (DNNs) has resulted in significant commercial adoption and deployment of the deep learning models for real-world applications. However, the superior accuracy of DNNs comes at the cost of (i) unsustainable computing demand for both model training and deployment, (ii) highly inefficient learning from massive amount of labeled training data, and (iii) non-robust and uninterpretable models with expensive manual model design and fine-tuning process.  Moreover, real world deployment environments can be highly uncertain and non-stationary, which often requires collecting new training data and retraining models when the environment/task changes. These inefficiencies and limitations of deep learning based solutions have tremendous consequences on cost, environment and society impact.

While pursuing DNN accuracy, we should also address the inefficiencies of current AI systems and their impact on our economy, society, and environment. In this talk, we will first discuss current deep learning limitations in terms of data, algorithms, and computation. Then, we will go through some early work on domain adaptation, autoML and inference optimization to address these challenges. Finally, a few future directions on making AI sustainable will be presented.