Home // ICSEA 2021, The Sixteenth International Conference on Software Engineering Advances // View article

Portable Fast Platform-Aware Neural Architecture Search for Edge/Mobile Computing AI Applications

Authors:
Kuo-Teng Ding
Hui-Shan Chen
Yi-Lun Pan
Hung-Hsin Chen
Yuan-Ching Lin
Shih-Hao Hung

Keywords: Portable, Neural Architecture Search, Performance Prediction

Abstract:
The recent rise and progress of neural network-based artificial intelligence are obvious, which we have settled up many traditional hard problems by deep learning. But the problem in neural network deployment on diverse hardware platforms is met, which needs lots of computational capability and time to ”try out” the best architecture tipping the balance of model accuracy and execution latency. The proposed Portable Fast Platform-Aware Neural Architecture Search (PFP-NAS) system allows users to use the trained neural network model easily without considering the hardware architecture of the edge/mobile computing on the client-side. The portable neural architecture search device shrinks the data center and converts it to allow users to utilize it on-demand and dynamically. This just-in-time, secure, and portable neural architecture search method is mainly based on the platform-aware client-side and applying the neural network model trained on the data center. Another key thing to remember is that users use the expandable modules of this device-Performance Prediction Module and Client Requirement-oriented Module: ie Accuracy, Latency, Throughput-FLOPs, MAC, Cost, etc., and then the device can detect hardware architectures such as USB/TPU/GPU/FPGA on edge/mobile computing. When detected, the device will send signals to connect the data center and drive the trained model in the data center for the corresponding hardware architecture. The proposed technique has the following characteristics: a. Only a boot medium is needed to detect and determine the hardware and then get the most suitable neural network from the server; b. Provide performance prediction module and client requirement-oriented module; c. Automatically match the model and the corresponding hardware architecture; d. Designed with modular scalability, and there is no need to configure any settings on the client-side. Consequently, the proposed framework achieves a portable data center.

Pages: 98 to 105

Copyright: Copyright (c) IARIA, 2021

Publication date: October 3, 2021

Published in: conference

ISSN: 2308-4235

ISBN: 978-1-61208-894-5

Location: Barcelona, Spain

Dates: from October 3, 2021 to October 7, 2021