Is machine learning energy-efficient enough to be used in mobile fashion applications?

Pintea, Paul (2022) Is machine learning energy-efficient enough to be used in mobile fashion applications? Bachelor's Thesis, Artificial Intelligence.

Preview	Text BachelorThesis_PaulPintea_S3593673.pdf Download (1MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (126kB)

Abstract

This report aims to analyze the energy efficiency of readily available mobile neural networks: MobileNet-V2, EfficientNet-lite0 and ResNet-50. MobileNet-V2 represents a lightweight neural network created for efficiency. EfficientNet-lite0 is a neural network built on a modern architecture that uses active scaling. ResNet-50 is a powerful neural network with 50 layers and high computational costs. All neural networks have been retrained on a modified version of the Fashion Product Images Dataset (Aggarwal, 2019). Testing focused on the energy consumption of the neural networks under an image classification task. The task was performed in a custom fashion-oriented application. Results show that MobileNet-V2 is the most energy-efficient, followed by EfficientNet-lite0 and then ResNet-50. A mean CPU usage of 26.5 to 28% was recorded for all three networks, while running on a Samsung A8 (2018). This represents a 12 to 13% increase from the simple version of the application. From the computational costs, ResNet-50 was expected to perform 10 times slower than the other two neural networks. This was not the case, from testing ResNet-50 ran 2 times longer, not 10 times, compared to the other two neural networks. This might be caused by the way multi-threaded mobile architectures work. Less demanding tasks occupy fewer threads and are not optimized as well, whilst more demanding tasks are split across multiple threads to improve energy efficiency. Even so, all three NNs seem energy efficient enough for deployment in different mobile environments.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Lawrence, C.P. and Jaeger, H.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	01 Mar 2022 10:14
Last Modified:	09 Mar 2022 13:45
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/26554

Actions (login required)

View Item