Thesis supervisor: Róbert Fullér
Location of studies (in Hungarian): Széchenyi István University Abbreviation of location of studies: SZE
Description of the research topic:
Deep neural networks have demonstrated their power in many computer vision applications. Recurrent Neural Networks deal with sequence-to-sequence prediction. It can be difficult to learn long-distance dependencies, in order to adjust the input-to-hidden weights based on the first input, the error signal needs to travel backwards through this entire pathway in principle, this lets us train them using the deepest (gradient) descent method. However, one of the main problems with the deepest descent method is that we need to learn dependencies over long time windows, and the gradients can explode or vanish. We’ll discuss at the problem itself, i.e. why gradients explode or vanish changing the architecture to one where the gradients are stable. This research is expected to provide new insight into the effectiveness of the method to inspired by the prominent success of proportional-integral-derivative controller in automatic control, so propose a proportional-integral-derivative approach for accelerating deep network optimization. Improvement in the current modeling techniques will improve the reliability of the Recurrent Neural Networks during training process
Number of students who can be accepted: 1
Deadline for application: 2021-04-30
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).