Aristotle University Of Thessaloniki (AUTH)
Information processing utilizing low-rank and sparse representations
Abstract
dc:descriptionThis doctoral dissertation introduces a series of algorithms and problem-solving frameworks in the fields of machine learning and optimization theory, with applications in forensics, computer vision, signal processing, and natural language processing. The research explores both unimodal and multimodal techniques for feature computation and pattern classification. Initially, the problem of computing principal components subject to non-negativity constraints on the elements of the eigenvectors was studied. To solve this problem, Lagrange multiplier theory was employed. Subsequently, using Lagrange multipliers, an algorithm was designed to solve a wide range of optimization problems in the positive orthant with equality constraints. Next, considering the multimodal nature of data, cross-lingual document classification and mobile device classification were examined. For cross-lingual document classification, a method for processing multidimensional data was extended to compute text representations in a latent space constrained in the positive orthant, connecting multilingual information. Moreover, for the multimodal classification of mobile devices, a cost function was proposed that combines neural networks trained on both image and audio data. With the rapid spread of multimedia in the digital age, concerns about theirprovenance have intensified. To address risks related to the origin of multimedia content, an algorithm was proposed for estimating the electric network frequency. Electric network frequency estimation has applications in identifying the location and time at which a multimedia recording, such as an audio message from a mobile phone, was captured. Solving the aforementioned problems requires the application of optimization algorithms that aim to determine the optimal parameters yielding a local minimum of each problem’s cost function. In most deep learning problems, the cost function is non-convex, which can lead not only to local minima but also to saddle points, where the gradient vanishes without necessarily yielding an optimal solution. To avoid saddle points, an algorithm was proposed that extends a second-order optimization method by making it non-parametric, thereby enhancing its capability to solve problems with minimal parameter tuning.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Tsingalis, Ioannis
- Τσιγγαλής, Ιωάννης
Subjects
dc:subject × 16- Μείωση διάστασης
- Θεωρία βελτιστοποίησης
- Ετερογλωσσική μεταφορά μάθησης
- Επεξεργασία σήματος
- Βαθιά μάθηση
- Dimensionality reduction
- Optimization theory
- Cross-lingual transfer learning
- Signal processing
- Deep learning
- 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/60362