Publications-Proceedings

Article View/Open

Publication Export

Google ScholarTM

NCCU Library

Citation Infomation

Related Publications in TAIR

TitleResidual Net Aspect of Disk stitching-based Manifold Reconstruction Method
Creator蔡炎龍
Tsai, Yen-Lung;Lin, Tse-Yu
Contributor應數系
Date2024-01
Date Issued12-Mar-2025 11:17:21 (UTC+8)
SummaryManifold learning is a branch of nonlinear dimensionality reduction based on the assumption that data of interest lies in a low-dimensional manifold embedded in a higher-dimensional space. The goal of manifold reconstruction is to reconstruct a Riemannian manifold for Euclidean and non-Euclidean datasets. One differential geometry-based framework of manifold reconstruction proposed by Fefferman et al (2020) state a reconstruction process from both of theoretical and computational aspects. Algorithm of this type is called the disk stitching method since multiple tangent affine spaces will be glued smoothly. In this work, we aim to provide a new point of view to rephrase this process as a learning process of a residual neural network model. Each local affine orthogonal projection can be viewed as a residual block satisfying some functional equations. This idea offers a new insight to study the interdisciplinary relationship of differential geometry and deep learning in the future.
Relation2024 Joint Mathematics Meetings (JMM 2024), AMS (American Mathematical Society)
Typeconference
dc.contributor 應數系
dc.creator (作者) 蔡炎龍
dc.creator (作者) Tsai, Yen-Lung;Lin, Tse-Yu
dc.date (日期) 2024-01
dc.date.accessioned 12-Mar-2025 11:17:21 (UTC+8)-
dc.date.available 12-Mar-2025 11:17:21 (UTC+8)-
dc.date.issued (上傳時間) 12-Mar-2025 11:17:21 (UTC+8)-
dc.identifier.uri (URI) https://nccur.lib.nccu.edu.tw/handle/140.119/156219-
dc.description.abstract (摘要) Manifold learning is a branch of nonlinear dimensionality reduction based on the assumption that data of interest lies in a low-dimensional manifold embedded in a higher-dimensional space. The goal of manifold reconstruction is to reconstruct a Riemannian manifold for Euclidean and non-Euclidean datasets. One differential geometry-based framework of manifold reconstruction proposed by Fefferman et al (2020) state a reconstruction process from both of theoretical and computational aspects. Algorithm of this type is called the disk stitching method since multiple tangent affine spaces will be glued smoothly. In this work, we aim to provide a new point of view to rephrase this process as a learning process of a residual neural network model. Each local affine orthogonal projection can be viewed as a residual block satisfying some functional equations. This idea offers a new insight to study the interdisciplinary relationship of differential geometry and deep learning in the future.
dc.format.extent 128 bytes-
dc.format.mimetype text/html-
dc.relation (關聯) 2024 Joint Mathematics Meetings (JMM 2024), AMS (American Mathematical Society)
dc.title (題名) Residual Net Aspect of Disk stitching-based Manifold Reconstruction Method
dc.type (資料類型) conference