DC Field | Value | Language |
---|---|---|
dc.contributor.advisor | Koh, In-Gyu | - |
dc.contributor.advisor | 고인규 | - |
dc.contributor.author | Han, Sang-Hee | - |
dc.contributor.author | 한상희 | - |
dc.date.accessioned | 2011-12-14T07:55:13Z | - |
dc.date.available | 2011-12-14T07:55:13Z | - |
dc.date.issued | 1999 | - |
dc.identifier.uri | http://library.kaist.ac.kr/search/detail/view.do?bibCtrlNo=151558&flag=dissertation | - |
dc.identifier.uri | http://hdl.handle.net/10203/48496 | - |
dc.description | 학위논문(석사) - 한국과학기술원 : 물리학과, 1999.2, [ ii, 30 p. ] | - |
dc.description.abstract | As perceptrons, the generalization power of layered feed-forward neural network(MLP) on the variations in test data is important for practical use of the neural networks. For the analysis of these stabilities in the MLP, we changed the feed-forward of neural network as cascade of $φ^4$ field theory with corresponding field Lagrangian without losing any properties of the network. Using this well-known topological soliton solution of field theory model, the stability of neural network is interpreted as topological stability of kink-type soliton solutions of $φ^4$ field theory. An explicit example, a simple three-class recognition problem, is shown with their theoretical field Lagrangian for practical manner. | eng |
dc.language | eng | - |
dc.publisher | 한국과학기술원 | - |
dc.subject | 위상적솔리톤해 | - |
dc.subject | 안정성 | - |
dc.subject | Topological stability | - |
dc.subject | Topological soliton | - |
dc.subject | Field theory | - |
dc.subject | Neural network | - |
dc.subject | 신경망 | - |
dc.title | Stability analysis of feed-forward neural network using topological soliton | - |
dc.title.alternative | 위상솔리톤을 이용한 신경망의 안정성에 관한 연구 | - |
dc.type | Thesis(Master) | - |
dc.identifier.CNRN | 151558/325007 | - |
dc.description.department | 한국과학기술원 : 물리학과, | - |
dc.identifier.uid | 000973729 | - |
dc.contributor.localauthor | Han, Sang-Hee | - |
dc.contributor.localauthor | 한상희 | - |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.