SUGIURA Riku, a second-year master’s student of the Mathematical Sciences Program, the Graduate School of Advanced Mathematical Sciences, received the FIT Incentive Award at the 23rd Information Science and Technology Forum (FIT2024).
This award is given to the best presentation in each session of the general presentations, chosen by the Chair at the Information Science and Technology Forum (FIT), a joint conference of the Information Processing Society of Japan, the IEICE Information and Systems Society, and the Human Communication Group.
Presentation title: Preventing a decline in the quality of learning on mixed datasets that include generated images using a descending-type decaying cyclic learning rate scheduler
Presentation summary: With the recent development of generative models, there are cases in which a mixed data set is used as a learning data set in machine learning, where a part of the real data is replaced with generated data. In this study, we introduced a descending-type decaying periodic scheduler as a measure against the decline in learning accuracy that occurs in such cases. The results of the verification using the mixed data set showed that the introduction method has a certain degree of effectiveness in preventing a decline in the quality of the generated images.
