Zobrazit v EDS

Poster Abstract: a ChainSGD-reduce Approach to Mobile Deep Learning for Personal Mobile Sensing

Uloženo v:
Podrobná bibliografie
Název: Poster Abstract: a ChainSGD-reduce Approach to Mobile Deep Learning for Personal Mobile Sensing
Autoři: Yu Zhang, Tao Gu, Xi Zhang
Rok vydání: 2020
Témata: Machine learning not elsewhere classified, Distribute computing, Mobile deep learning, Neural networks, Reinforcement learning, Resource allocation
Popis: MDLdroid is a novel decentralized mobile deep learning framework, which enables resource-aware on-device collaborative learning for personal mobile sensing applications. To address resource limitation, MDLdroid uses a chain-directed Synchronous Stochastic Gradient Descent (ChainSGD-reduce) approach to effectively reduce overhead among multiple devices. In addition, MDLdroid includes an agent-based multi-goal reinforcement learning mechanism to balance resources in a fair and efficient manner. Real-world experiments demonstrate that our model training on off-the-shelf mobile devices achieves 2× to 3.5× faster than single-device training, and 1.5× faster than the master-slave approach.
Druh dokumentu: conference object
Jazyk: unknown
Relation: 10779/rmit.27589056.v1
Dostupnost: https://figshare.com/articles/conference_contribution/Poster_Abstract_a_ChainSGD-reduce_Approach_to_Mobile_Deep_Learning_for_Personal_Mobile_Sensing/27589056
Rights: All rights reserved
Přístupové číslo: edsbas.81C7BCC4
Databáze: BASE
Popis
Abstrakt:MDLdroid is a novel decentralized mobile deep learning framework, which enables resource-aware on-device collaborative learning for personal mobile sensing applications. To address resource limitation, MDLdroid uses a chain-directed Synchronous Stochastic Gradient Descent (ChainSGD-reduce) approach to effectively reduce overhead among multiple devices. In addition, MDLdroid includes an agent-based multi-goal reinforcement learning mechanism to balance resources in a fair and efficient manner. Real-world experiments demonstrate that our model training on off-the-shelf mobile devices achieves 2× to 3.5× faster than single-device training, and 1.5× faster than the master-slave approach.