Thesis supervisor: Bálint Gyires-Tóth
Location of studies (in Hungarian): Department of Telecommunications and Media Informatics Abbreviation of location of studies: TMIT
Description of the research topic:
Deep learning has revolutionized many research areas, like speech recognition and synthesis, image classification and segmentation and natural language understanding. The modeling performance of deep learning systems significantly surpasses earlier machine learning methods due to their ability to extract and to learn various abstractions of huge amount of input data. However, in case of few training examples, their performance remains typically poor, and new data leads to catastrophic forgetting of previous knowledge. Generally, deep neural networks have ‘black-box’ type behavior – the reasoning and inference rule of deep learning systems stays unknown throughout the whole process. In contrast, humans are capable of learning complex processes based on only few samples by transferring the knowledge of similar domains. Humans can also explain the logic behind their reasoning and inference.
The goal of the current research is to create novel algorithms that help to understand and extract the reasoning and consider knowledge representation in deep neural networks. Thus, fewer learning examples will be enough to train neural networks, and the training process may be aided by previous knowledge. The effectiveness of the elaborated method must be proven at least in one application scenario. Such an application scenario can be (1) sequential and temporal data modeling, (2) natural language processing, (3) speech technologies, etc.
The research can be performed in English and Hungarian. For training the models public and private databases and high-performance GPUs are available.
The possible research tasks of the Ph.D. student are the following:
• Overview the related scientific papers, including basic deep neural networks elements and novel results in knowledge extraction of neural networks.
• Design and implement a baseline system for modeling training samples of similar domains separately.
• Elaborate novel algorithms to extract the knowledge from trained neural networks and transfer it to a new network from the similar domain.
• Demonstrate the effectiveness of the results at least in one application scenario.
• Objective and subjective evaluation.
Required language skills: english Further requirements: Programming and mathematical skills.
Number of students who can be accepted: 1
Deadline for application: 2019-01-07
2024. IV. 17. ODT ülés Az ODT következő ülésére 2024. június 14-én, pénteken 10.00 órakor kerül sor a Semmelweis Egyetem Szenátusi termében (Bp. Üllői út 26. I. emelet).
Login
