Back to search

Πανεπιστήμιο Μακεδονίας

Parallel and distributed methods for autonomous design of artificial neural networks

Abstract

dc:description

In recent years, neural networks in the form of deep learning have re-ignited a widespread interest in artificial intelligence. The applications that leverage deep learning have automated several conceptionally easy tasks that are hard to express algorithmically. Examples include identifying objects in an image, counting the number of occurrences for the same object, autonomous driving and accurate speech-to-text. While deep learning enables the automation of such tasks without explicitly instructing the machine, intricate neural architectures are employed. These architectures require several person-hours and computational resources to design them. Furthermore, expertise both in neural networks and in the specific field of application are pre-requisites.Neural architecture search aims to automate the process of designing the networks by several different approaches, including metaheuristics, reinforcement learning, and differentiable methods. This thesis employs and studies parallel and distributed neural architecture search algorithms. The efficiency and effectiveness of the implemented algorithms are studied and speed-up techniques that are widely adopted but have not been studied in-depth before. Furthermore, under the scope of the thesis, an open-source library aimed at distributed neural architecture search has been developed. Finally, the thesis contributes to the field of neural architecture search by verifying the effectiveness and showing the limitations of methods such as reduced epoch training, distributed evolutionary search methods, distributed differentiable methods, and employing neural architecture search methods for graph convolutional networks.

Degree

thesis:*
Grantor dc:publisher
Πανεπιστήμιο Μακεδονίας
Year dc:date
2022

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Kyriakides, Georgios
  • Κυριακίδης, Γεώργιος

Subjects

dc:subject × 12

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/51944

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Kyriakides, Georgios; Κυριακίδης, Γεώργιος. Parallel and distributed methods for autonomous design of artificial neural networks. Πανεπιστήμιο Μακεδονίας, 2022. http://hdl.handle.net/10442/hedi/51944