Aristotle University Of Thessaloniki (AUTH)
Supervised and unsupervised deep learning methodologies
Abstract
dc:descriptionIn recent years, Deep Learning (DL) methods have been excelling in various tasks, including classification, clustering, and regression among others. Part of their success is the introduction of successively deeper architectures, which have been shown to increase the performance in tasks like these, at the cost of significantly increased computational complexity. In this PhD thesis, we investigated various DL methods which fall under the Supervised and Unsupervised Learning cases, with the goal of increasing their performance whilst maintaining their computational cost at low levels. The choice to keep computational costs low was motivated, for one, by the increasing popularity of DL methods, as applications like object detection and tracking, among others, have found their way into almost all modern mobile devices. Furthermore, the advent of drones, and other robots, which currently can only take on small, embedded devices, also influenced the algorithms investigated in this thesis. First, a general framework for discriminant autoencoding was proposed, to tackle the problems of classification and clustering, using lightweight neural networks. Towards this goal, Autoencoders (AEs) were studied, and various discriminative criteria were inserted into their learning objective, under the general Graph Embedding Framework. Both supervised and unsupervised criteria were proposed, and the resulting representations obtained by the proposed Discriminant Autoencoders were evaluated on various benchmarks, showcasing their ability to generalize and apply the criteria to unseen samples at no added cost. Next, the problem of visual object tracking was tackled, with the goal of using the proposed methods on embedded devices. Specifically, a generic framework of parallel object detection and tracking was proposed, which supports the use of any object tracker and significantly boosts its performance at virtually no additional cost. In the same task, a different approach was followed to develop a tracker based on a deep convolutional feature histogram architecture. Histograms have been used in this task in earlier years, and the combination of histograms with convolutional features, as well as that of features from different levels of the network, was proven experimentally to be superior to the baseline methods. Next, unsupervised autoencoders were once again investigated, this time in the field of Astrophysics, and specifically in gravitational wave surrogate modelling. The use of AEs in this task helped us uncover a latent spiral relationship between the surrogate coefficients and the corresponding independent variables, which we were able to formulate and incorporate into lightweight neural networks which learned to regress the coefficients more faithfully. Finally, returning to the classification task, the use of label embedding methods was proposed, to tackle the issue of inherent ambiguity present in classification datasets which use binary labels to encode the data classes. A general-class and an instance-specific variant were proposed and shown to ease the training process and lead to better generalization.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2021
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Nousi, Paraskevi
- Νούση, Παρασκευή
Subjects
dc:subject × 14- Επιβλεπόμενη μάθηση
- Μη επιβλεπόμενη μάθηση
- Βαθιά μάθηση
- Νευρωνικά δίκτυα
- Supervised learning
- Unsupervised learning
- Deep learning
- Neural networks
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/51392