Edge computing techniques for classification based multi-task incremental learning

Dube, Swaraj Sunilkumar (2023) Edge computing techniques for classification based multi-task incremental learning. PhD thesis, University of Nottingham.

[img] PDF (Thesis - as examined) - Repository staff only - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
Available under Licence Creative Commons Attribution.
Download (3MB)

Abstract

The number of Internet of Things (IoT) edge devices are exponentially on the rise that have both computational capabilities and internet connectivity. IoT devices are the ones that collect massive amounts of data which can then be analyzed and can also be used to train machine learning models for various applications. However, edge devices, generally do not possess the same resources as the cloud in terms of computing power and storage. Such limitations of edge devices often lead to relying largely on the cloud for further data processing. Consequently, large transmission costs occur with respect to the data size and number of edge devices connected to the cloud. Furthermore, real-world data always evolves over time, hence machine learning needs to be learning continuously to adapt to such evolving data, resulting in large data transmission over time.

The main focus of this research is to reduce data transmission costs between edge devices and the cloud in an incremental learning scenario while retaining deep neural network model performance and to reduce the energy consumption of the learning process on multiple edge devices.

Firstly, a novel pre-training data sampling algorithm is presented for edge-cloud systems whereby this algorithm can automatically select a subsample from a training distribution right at the edge device before training even starts while maintaining the incremental learning performance of a Convolutional Neural Network (CNN) in terms of its classification accuracy. This means that if this data sampling technique is executed on edge devices, it will not require the transmission of all data samples to the cloud. This can save both the transmission cost and the training cost on the cloud while the model performance is largely retained. Results show that after performing the developed pre-training data sampling technique, the classification performance of Convolutional Neural Network (CNN) models in the context of incremental learning remains within 3% while the transmission costs are reduced by up to 18%.

Secondly, a novel method to extract only the important weights of a fully connected neural network is also presented in this report. The method can also further reduce the transmission costs between the edge devices and the cloud. Experiments show that not all the weights learned by a neural network in the cloud are useful. So, it is actually not required to send all the learned weights back to the edge devices. When only the useful weights are transmitted back to the edge devices, this will be beneficial especially in the case of incremental learning where the size of a neural network can continuously grow. Results show that by implementing the proposed extraction method, the transmission costs can be reduced by at least 35%. The proposed extraction method can also extract useful weights/parameters faster than the baseline by a factor of at least 1.9.

Finally, this research also discovers that when multiple edge devices are participating in carrying out multi-task learning, the overall energy consumption across all devices is highly dependent on the edge device training order i.e., which edge device should start the training and which edge device should end the training. This is due to the heterogeneity of the edge devices i.e., some edge devices can be much more computationally powerful than the other edge devices. An introductory algorithm that can determine the edge device training order to minimize the end-to-end energy consumption across all the participating edge devices is developed. Results show that by following the introductory algorithm in choosing the correct order of edge devices for training, the overall end-to-end energy consumption can be reduced by a factor of at least 1.6.

Item Type: Thesis (University of Nottingham only) (PhD)
Supervisors: Nugroho, Hermawan
Keywords: data transmission costs, Internet of Things devices, edge devices, edge-cloud system
Subjects: T Technology > TK Electrical engineering. Electronics Nuclear engineering
Faculties/Schools: University of Nottingham, Malaysia > Faculty of Science and Engineering — Engineering > Department of Electrical and Electronic Engineering
Item ID: 69708
Depositing User: DUBE, Swaraj
Date Deposited: 18 Feb 2023 04:40
Last Modified: 18 Feb 2023 04:40
URI: https://eprints.nottingham.ac.uk/id/eprint/69708

Actions (Archive Staff Only)

Edit View Edit View