Repository logo
 

Optimization of AI models as the Main Component in Prospective Edge Intelligence Applications

dc.contributor.committeeMemberMohamed, Ebrahim Bedeer
dc.contributor.committeeMemberWahid, Khan
dc.contributor.committeeMemberBui, Francis
dc.contributor.committeeMemberStavness, Ian
dc.contributor.committeeMemberKo, Seok-Bum
dc.creatorTorres Tello, Julio Wladimir
dc.creator.orcid0000-0001-9694-6578
dc.date.accessioned2022-08-10T15:38:21Z
dc.date.available2022-08-10T15:38:21Z
dc.date.copyright2022
dc.date.created2022-08
dc.date.issued2022-08-10
dc.date.submittedAugust 2022
dc.date.updated2022-08-10T15:38:22Z
dc.description.abstractArtificial Intelligence (AI) is a successful paradigm with application in many fields; however, there can be some challenging scenarios like the deployment of AI models in remote locations or with limited connectivity, possibly needing to perform inference closer to where data is collected. A potential solution is the study of ways to optimize AI models, for deployment of intelligent algorithms closer to the edge. This thesis focuses on applications of AI that need to have characteristics that make them suitable for implementation on portable devices (e.g., aeroponics container, drone, mobile robot); thus, the development and implementation of custom models, and their optimization (i.e., reduction in size and inference time) is of upmost importance and the main goal of this dissertation. For this task, a number of options have been explored, including developing techniques to select relevant features from the samples that the model analyzes, and pruning and quantization. Therefore, this thesis proposes a scheme for the development, implementation, and optimization of custom AI models used mainly in agriculture, so that they have the desired characteristics that are needed for their deployment in edge devices. This main goal is fulfilled by implementing a number of sequential steps that include the validation of the hypothesis that there is at least an AI model capable of generating useful predictions for the applications being studied, the exploration and implementation of an approach for their optimization, and their final implementation in hardware of limited resources. The main contributions of this thesis are on the general workflow for optimization of custom models, as well as in the proposed scheme for feature selection based on model interpretability approaches. This carries most of the novelty of the thesis, since we have not found previous implementations of these ideas, at least in the field under study. This optimization is mainly based on a feature selection approach, but finally complemented with pruning and quantization. The implementation of some of these models on an edge-like device, demonstrates the feasibility of this approach. Finally, although this thesis tries to be a self-contained work, encompassing all the aspects required to go from AI model design to implementation on an edge device, there are some aspects that could be further studied, analyzed, and improved. Furthermore, it is almost impossible to keep the pace with all the new developments in the fields of AI, edge computing, hardware and software tools, etc. which opens the field for new discussions and proposals. This work tries to fill some gaps and to propose some ideas that hopefully will be useful for future researchers in the development of new technologies and solutions.
dc.format.mimetypeapplication/pdf
dc.identifier.urihttps://hdl.handle.net/10388/14092
dc.language.isoen
dc.subjectartificial intelligence
dc.subjectedge intelligence
dc.subjectfeature selection
dc.subjectinterpretability
dc.subjectsmart agriculture
dc.subjectremote sensing
dc.titleOptimization of AI models as the Main Component in Prospective Edge Intelligence Applications
dc.typeThesis
dc.type.materialtext
thesis.degree.departmentElectrical and Computer Engineering
thesis.degree.disciplineElectrical Engineering
thesis.degree.grantorUniversity of Saskatchewan
thesis.degree.levelDoctoral
thesis.degree.nameDoctor of Philosophy (Ph.D.)

Files

Original bundle
Now showing 1 - 1 of 1
Loading...
Thumbnail Image
Name:
TORRESTELLO-DISSERTATION-2022.pdf
Size:
12.54 MB
Format:
Adobe Portable Document Format
License bundle
Now showing 1 - 1 of 1
No Thumbnail Available
Name:
LICENSE.txt
Size:
2.27 KB
Format:
Plain Text
Description: