< BK21+ BEST Seminar Series Announcement> 

Time and Date : 16:00 ~ 18:00 Monday 10/31/2018

Place : D403, Engineering Building #4

Title : Mobile Edge Intelligence at Scale

Abstract:

In just a few years, breakthroughs in machine learning (ML) have transformed how computational models perform a wide variety of applications such as face recognition, autonomous driving, and language translation. Expanding its territory to mission-critical wireless applications makes the impossible become possible, e.g., supporting a massive number of devices with low latency and high reliability guarantees. This calls for a paradigm shift from classical ML to edge ML. To abide by the application’s low latency constraint, the inference in edge ML is operated at the network edges such as base stations (BSs) and mobile devices. Furthermore, the training process is performed locally at the edge side, and is collectively updated over the wirelessly connected edge devices, while not directly exchanging private data samples. Preserving data privacy enables the access to a huge amount of the user-generated training dataset, thereby improving the inference reliability. Unfortunately, traditional deep models and training algorithms exert severe demands on the energy/memory/communication resources of local devices, thus limiting their adoption for edge ML. Motivated by this, in this paper we explore the building blocks of edge ML, and propose its suitable theoretical and technical enablers, followed by several case studies demonstrating the effectiveness of our proposed approaches.

Presenter: Jihong Park, Ph.D. / Oulu University, Finland

Host: Prof. Kim, Seonglyun, Yonsei EEE