Georgia Institute of Technology
Statistical viewpoints on network model, PDE Identification, low-rank matrix estimation and deep learning
Abstract
dc:description.abstractThe phenomenal advancements in modern computational infrastructure enable the massive amounts of data acquisition in high-dimensional feature space possible. To put it more specific, the largest datasets available in the industry which often involve up to billions of samples and millions of features. The nature of datasets arising in modern science and engineering are sometimes even larger, often with the dimension of the same order as, or possibly even larger than, the sample size. The cornerstone of modern statistics and machine learning has been a precise characterization of how well we can estimate the objects of interests under these huge high-dimensional datasets. While it remains impossible to consistently estimate in such a high-dimensional regime in general, a large body of research has investigated various structural assumptions under which statistical recovery is possible even in these seemingly ill-posed scenarios. Examples include a large line of works on sparsity, low-rank assumptions and more abstract generalizations of these. These structural assumptions on signals are often realized through specially designed norms; i.e., for inducing sparsity of either vector or matrix, entry-wise L1-norm is used; for inducing low-rank matrix, nuclear norm is used. Not only in parametric, but in non-parametric models, high-dimensional dataset is common in real world applications. A deep neural network, one of the most successful models in modern machine learning in various tasks, is a primary example of non-parametric model for function estimations. Tasks such as image classification or speech recognition often require a dataset in high-dimensional space. For the accurate function estimation avoiding the commonly known curse of dimensionality phenomena, some special structural assumptions on regression functions are imposed. Under some specific structural assumptions imposed on problems, the main emphasis in this thesis proposal is on exploring how various regularizing penalties can be utilized for estimating parameters and functions in parametric and non-parametric statistical problems. Specifically, our main focus will be the problems in network science, PDE identification, and neural network.
Degree
thesis:*- Level thesis:degree_level
- Doctoral
- Department dc:contributor.department
- Industrial and Systems Engineering
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2022
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Suh, Namjoon
- Advisors dc:contributor.advisor
-
- Mei, Yajun
- Huo, Xiaoming
- Committee members dc:contributor.committeemember
-
- Kang, Sung ha
- Zhilova, Mayya
- Zhou, Ding-Xuan
Subjects
dc:subject × 5Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- http://hdl.handle.net/1853/70140