Supervised machine learning based one-shot learning at image classes

OData support
Supervisor:
Dr. Szűcs Gábor
Department of Telecommunications and Media Informatics

The evolution of artificial intelligence has increased dramatically over the last decade. Deep neural networks have achieved similar results to humans in image classification tasks. Their success is due to the number of training samples. But there are problems where we do not have a lot of training samples, just 1 or very few. In order to succeed with neural networks, we need to use new techniques in these circumstances.

In my thesis work, I present the theoretical background and known techniques of the topic (so-called one-shot learning). Choosing the Siamese network solution, I composed my scheme and implementation for the problem which will be presented in detail. I used the Omniglot dataset to analyze the effectiveness of my algorithm. After evaluating the results of the network, I suggested further development improvements. When I convinced the Siamese network usability I looked for the question of how one-shot algorithms can be used to detect similarities between siblings. To answer the question, I worked with the Siblings Database.

Downloads

Please sign in to download the files of this thesis.