Abstract
dc:descriptionRecent advances in Deep Learning (DL) have provided powerful tools for various data analysis tasks. However, DL methods suffer from high complexity, hindering their successful application when limited computational resources are available, while combining them with traditional machine learning methods is not always a straightforward task, further limiting their flexibility. The primary focus of this Ph.D. thesis is to develop DL methods that can tackle a wide range of different data analysis tasks (classification, clustering, regression, etc.) using any kind of data (image, video, text, time series), while overcoming the aforementioned limitations. First, a neural generalization of the Bag-of-Features model is proposed and combined with various feature extractors, including Deep Convolutional Neural Networks, increasing the accuracy of the networks and providing much better invariance to distribution shifts, while greatly reducing the number of parameters of the model over the competitive methods. Then, a representation learning method, that produces retrieval-oriented representations by adhering to the cluster hypothesis, is proposed, allowing for significantly improving the retrieval performance. Furthermore, an efficient and flexible knowledge transfer method, that can effectively distill the knowledge encoded in a large and complex neural network into a smaller and faster one, is proposed and evaluated under various setups. Next, the problem of dimensionality reduction is formulated as a knowledge transfer problem, where the knowledge is to be transferred from an appropriately defined Probability Density Function (PDF) into a machine learning model and the aforementioned knowledge transfer method is further extended providing a generic dimensionality method. The proposed framework is also combined with deep learning models to derive a manifold-based representation learning method that can provide representations optimized for clustering tasks. Finally, an open-source library that implements the developed dimensionality reduction framework was developed and it is presented in this thesis, along with a temporal averaging method that can improve the convergence of stochastic optimization methods.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2018
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Passalis, Nikolaos
- Πασσαλής, Νικόλαος
Subjects
dc:subject × 10Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/44841