Item – Theses Canada

​
OCLC number
1006674320
Author
Sun, Xiang,1968-
Title
The Lasso and its implementation for neural networks.
Degree
Ph. D. -- University of Toronto, 1999
Publisher
Ottawa : National Library of Canada = Bibliothèque nationale du Canada, [2000]
Description
2 microfiches
Notes
Includes bibliographical references.
Abstract
Choosing the architecture of a neural network is one of the most important problems in making neural networks practically useful. How many hidden units are needed, should shrinkage be used, and if so how much? This thesis proposes Lasso neural network models and algorithms for fitting models based on feed-forward neural networks. Lasso neural network models select the architecture of the neural network in the process of estimating its parameters. These models employ both the Lasso style constraint and the idea of hyperparameters of the ARD model, and combine them in a manner that preserves the Lasso's advantage of shrinking some parameters to exactly zero and the hyperparameter's advantage of controlling the magnitudes of predictors and/or hidden units. We also prove that in linear models, the Lasso solution function is a linear spline in terms of the shrinkage parameter, and derive the closed forms of the sub-functions of this spline. Accordingly, we propose a quasi-analytic algorithm for solving the Lasso solutions in linear models. Since the Lasso solutions are equivalent to their Ridge- modified version, the more accurate Ridge-modified GCV scores for Lasso solutions can be obtained by the quasi-analytic algorithm.
ISBN
0612457958
9780612457959