DC Field | Value | Language |
---|---|---|
dc.contributor.advisor | Yoon, Kuk-Jin | - |
dc.contributor.advisor | 윤국진 | - |
dc.contributor.author | Jeong, Wooseong | - |
dc.date.accessioned | 2023-06-22T19:30:43Z | - |
dc.date.available | 2023-06-22T19:30:43Z | - |
dc.date.issued | 2023 | - |
dc.identifier.uri | http://library.kaist.ac.kr/search/detail/view.do?bibCtrlNo=1032298&flag=dissertation | en_US |
dc.identifier.uri | http://hdl.handle.net/10203/308096 | - |
dc.description | 학위논문(석사) - 한국과학기술원 : 기계공학과, 2023.2,[vi, 48 p. :] | - |
dc.description.abstract | Although suppressing negative transfer between tasks has been a critical challenge for multi-task learning, previous approaches have dealt with multi-task architecture and optimization strategies separately for the purpose. Instead, I propose connection strength-based optimization with progressive multi-modal feature exchange as a combined method for reducing task interference by (i) conserving each task’s feature space in a shared network, and (ii) facilitating inter-task information flow in feature level. I reinterpret the connection strength, a well-known concept in network compression, to determine which channel of the shared convolutional layer has a dominant influence on each task. Based on connection strength, the proposed optimization method projects each task’s gradient to prevent a specific task from exerting a dominant influence on the entire network by intruding on other tasks’ space. Then, these conserved features are progressively forwarded and mixed in a stage-by-stage manner from a shared single-tasking backbone, so that the network fully utilizes inter-task information by exchanging task-specific features. I propose an integrated method that conserves tasks feature space, enabling hard parameter sharing with minimized task interference and allowing it to be used for multi-modal feature exchange. Experiments demonstrate the validity of proposed methods on several dense-prediction tasks by achieving state-of-the- art performances. | - |
dc.language | eng | - |
dc.publisher | 한국과학기술원 | - |
dc.subject | 다중작업 학습▼a연결 강도▼a기울기 벡터 투사▼a다중 양식▼a특징 융합 | - |
dc.subject | Multi-task learning▼aconnection strength▼agradient projection▼amulti-modal▼afeature fusing | - |
dc.title | Connection strength-based optimization with progressive multi-modal feature exchange for multi-task learning | - |
dc.title.alternative | 다중 작업 학습을 위한 연결강도 기반 최적화와 점진적 다중 형식 특징 교환 | - |
dc.type | Thesis(Master) | - |
dc.identifier.CNRN | 325007 | - |
dc.description.department | 한국과학기술원 :기계공학과, | - |
dc.contributor.alternativeauthor | 정우성 | - |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.