Deep Learning on the mobile edge Georg Eickelpasch advised by - PowerPoint PPT Presentation

Chair of Network Architectures and Services Department of Informatics Technical University of Munich Deep Learning on the mobile edge Georg Eickelpasch advised by Marton Kajo Thursday 10 th October, 2019 Chair of Network Architectures and Services Department of Informatics Technical University of Munich

Structure • Introduction • Use cases • Scheduling strategies G. Eickelpasch — DL on the edge 2

Introduction Problem of mobile Deep Learning • Offloading to cloud • Bottleneck bandwidth • Unstable mobile connection • Compute on device • Compute using the edge G. Eickelpasch — DL on the edge 3

Introduction Terminology of Edge 1. Edge as mobile device -> offload from 2. Edge as local server -> offload to Possible layers in an edge system G. Eickelpasch — DL on the edge 4

Use cases Speech • Speech recognition • Natural Language Processing • Real-time processing • Robustness • Complex Data G. Eickelpasch — DL on the edge 5

Use cases Computer Vision • Object identification • Continuous Vision • Real-time processing • Mobile context • Large files G. Eickelpasch — DL on the edge 6

Use cases E-Health • Health Tracking • Diseas identification • Highly sensitive data • Requiered robustness • Heterogeneous data G. Eickelpasch — DL on the edge 7

Scheduling Preprocessing Strategy Simplified overview for NeuroSurgeon algorithm G. Eickelpasch — DL on the edge 8

Scheduling Preprocessing Strategy G. Eickelpasch — DL on the edge 9

Scheduling Deadline Strategy Simplified overview for Edgent algorithm G. Eickelpasch — DL on the edge 10

Scheduling Deadline Strategy G. Eickelpasch — DL on the edge 11

Conclusion • NeuroSurgeon • only optimization • weak edge • Edgent • best result in given time • strong edge G. Eickelpasch — DL on the edge 12

Future work • Optimization of Edgent with NeuroSurgeon approach Preprocessing on the device as well as early exit G. Eickelpasch — DL on the edge 13

Discussion questions • Do you think Edge Computing will be the dominant solution in the future? G. Eickelpasch — DL on the edge 14