Πανεπιστήμιο Μακεδονίας
Decentralized deep neural network training via distributed ledger technology
Abstract
dc:descriptionArtificial Intelligence has been proven to hugely benefit from big data and enormous computer power. This fact gives a great head start to the few large companies that possess such infrastructure, so they tend to dominate the AI space. A desirable alternative could be the creation of prototypes that: 1) They no longer rely on large private data centers. 2) They are not based on just a decentralized model, but on a fully distributed one, meaning that they have no need for a central coordinator. Following the above directives, we could develop networks that: 1) They can approach (or even exceed) the total computing power of a big data center, by utilizing a large number of nodes. 2) They have no "single point of failure", since all participants have equal roles, meaning that the network will keep working even if a portion of it gets offline. 3) They are democratized, since all nodes have the exact same access to the produced outcome. So far almost all of the implementations regarding neural network training, focus mainly on performance issues. Although there has been significant research on distributed algorithms, there is not enough research on a model that has no need of a coordinating server. This thesis utilizes new peer-to-peer topologies based on cryptographic verification (Distributed Ledger Technology - DLT), attempting to diffuse the server role to the whole network. During our research we tested various configurations of the distributed model, in order to evaluate their impact on both the training speed and the quality of the produced model. The result is a novel and ready-to-use ecosystem, for a democratized method of training deep neural networks.
Degree
thesis:*- Grantor dc:publisher
- Πανεπιστήμιο Μακεδονίας
- Year dc:date
- 2022
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Nikolaidis, Spyridon
- Νικολαΐδης, Σπυρίδων
Subjects
dc:subject × 18- Τεχνητή νοημοσύνη
- Μηχανική μάθηση
- Νευρωνικά δίκτυα
- Διαδίκτυο των πραγμάτων
- Αποκεντρωμένες αρχιτεκτονικές
- Artificial intelligence
- Machine learning
- Artificial neural networks
- Distributed ledger
- Blockchain
- Internet of things (IoT)
- Decentralized architectures
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Υπολογιστές, Υλικό (hardware) και Αρχιτεκτονική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Computer science, Hardware and Architecture
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/51333