Korkmaz, Ezgi;
(2024)
Principled Analysis of
Machine Learning Paradigms.
Doctoral thesis (Ph.D), UCL (University College London).
Preview |
Text
EzgiKorkmazPhDThesis.pdf - Published Version Download (18MB) | Preview |
Abstract
Utilization of deep neural networks as function approximators to learn policies that can make sequential decisions in high-dimensional complex MDPs has led to striking progress for artificial intelligence and machine learning applications. More than a decade of extensive research has enabled us to build artificial intelligence that can transcend natural intelligence in highly complex tasks, yet on the other hand our lack of foundational understanding on the knowledge the artificial intelligences form and the fundamental limitations behind it stands in stark contrast to the exhibited capabilities of these models. In this thesis, first I introduce the concept of reasoning over decisions in artificial intelligence to develop understanding on what truly artificial intelligences learn, and further demonstrate that intelligent agents that reason over their decisions can identify and understand their limitations, capabilities and their knowledge of the world. In the second part of the thesis, I introduce a framework to establish a foundational scientific analysis on the functions learnt by artificial intelligence agents, the underlying composition of the decision boundaries and the structure of the loss landscape. Furthermore, I provide mathematical analysis on the limitations of reinforcement learning that explains the inevitability of high-sensitivity directions in high-dimensions. I conduct extensive empirical analysis in high dimensional complex MDPs and discover that high-sensitivity directions for deep reinforcement learning policies are shared across MDPs and further across algorithms. I further introduce the theoretical foundations establishing that artificial intelligence agents are intrinsically bounded by the geometry of high-dimensions. Both the theoretical analysis I introduce and the empirical analysis I conduct establish the principles of learning in high-dimensions, and the limitations behind it. In the final part of the thesis, I introduce a contradistinction analysis between natural and artificial intelligence. This juxtaposition demonstrates the intrinsic differences between how humans solve the same given tasks compared to intelligent agents. These intrinsic differences between artificial and natural intelligence provide further insights towards understanding how artificial intelligences perceive the world, and the limitations they have that bound their current capabilities to abstract and generalize to complex uncertain environments.
| Type: | Thesis (Doctoral) |
|---|---|
| Qualification: | Ph.D |
| Title: | Principled Analysis of Machine Learning Paradigms |
| Open access status: | An open access version is available from UCL Discovery |
| Language: | English |
| Additional information: | Copyright © The Author 2024. Original content in this thesis is licensed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) Licence (https://creativecommons.org/licenses/by-nc-nd/4.0/). Any third-party copyright material present remains the property of its respective owner(s) and is licensed under its existing terms. Access may initially be restricted at the author’s request. |
| UCL classification: | UCL UCL > Provost and Vice Provost Offices > UCL BEAMS UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Engineering Science > Dept of Computer Science |
| URI: | https://discovery.ucl.ac.uk/id/eprint/10198328 |
Archive Staff Only
 |
View Item |
