DC Field | Value | Language |
---|---|---|
dc.contributor.advisor | Park, Jinkyoo | - |
dc.contributor.advisor | 박진규 | - |
dc.contributor.author | Yu, Jihwan | - |
dc.date.accessioned | 2023-06-23T19:31:11Z | - |
dc.date.available | 2023-06-23T19:31:11Z | - |
dc.date.issued | 2023 | - |
dc.identifier.uri | http://library.kaist.ac.kr/search/detail/view.do?bibCtrlNo=1032748&flag=dissertation | en_US |
dc.identifier.uri | http://hdl.handle.net/10203/308792 | - |
dc.description | 학위논문(석사) - 한국과학기술원 : 산업및시스템공학과, 2023.2,[iii, 15 p. :] | - |
dc.description.abstract | Deep Learning typically learns the parameters of neural network models using large amounts of data of a given task. However, in recent years, researchers conducted a study to improve the performance of multi-task learning problems based on a small amount of data. Among them, optimization-based meta-learning is a method of learning parameters of new tasks using optimization-based fine-tuning from learned initial parameters. However, there is a lack of an analytic approach that considers the correlation between tasks from the perspective of game theory. In this study, we propose a task-specific optimization problem using joint constraints. We then present the generalized Nash equilibrium between tasks using the variational equilibrium and present a new meta-learning algorithm. Finally, in the simulations, the performance of the proposed model is compared with the existing optimization-based meta-learning model by the sinusoidal regression problem. | - |
dc.language | eng | - |
dc.publisher | 한국과학기술원 | - |
dc.subject | Meta-learning▼aGame theory▼aVariational equilibrium | - |
dc.subject | 메타 학습▼a게임 이론▼a변동 균형 | - |
dc.title | (A) game theoretic approach to model-agnostic meta-learning using variational equilibrium | - |
dc.title.alternative | 변동 균형을 이용한 일반적인 모델에 적용할 수 있는 메타 학습의 게임 이론적 접근법 | - |
dc.type | Thesis(Master) | - |
dc.identifier.CNRN | 325007 | - |
dc.description.department | 한국과학기술원 :산업및시스템공학과, | - |
dc.contributor.alternativeauthor | 유지환 | - |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.