Off-campus UMass Amherst users: To download campus access dissertations, please use the following link to log into our proxy server with your UMass Amherst user name and password.

Non-UMass Amherst users: Please talk to your librarian about requesting this dissertation through interlibrary loan.

Dissertations that have an embargo placed on them will not be available to anyone until the embargo expires.

Author ORCID Identifier

https://orcid.org/0000-0002-4286-4793

Document Type

Open Access Dissertation

Degree Name

Doctor of Philosophy (PhD)

Degree Program

Electrical and Computer Engineering

Year Degree Awarded

2019

Month Degree Awarded

May

First Advisor

Marco F. Duarte

Subject Categories

Other Computer Engineering | Signal Processing

Abstract

Today we are living in a world awash with data. Large volumes of data are acquired, analyzed and applied to tasks through machine learning algorithms in nearly every area of science, business, and industry. For example, medical scientists analyze the gene expression data from a single specimen to learn the underlying causes of disease (e.g. cancer) and choose the best treatment; retailers can know more about customers' shopping habits from retail data to adjust their business strategies to better appeal to customers; suppliers can enhance supply chain success through supply chain systems built on knowledge sharing. However, it is also reasonable to doubt whether all the genes make contributions to a disease; whether all the data obtained from existing customers can be applied to a new customer; whether all shared knowledge in the supply network is useful to a specific supply scenario. Therefore, it is crucial to sort through the massive information provided by data and keep what we really need. This process is referred to as information selection, which keeps the information that helps improve the performance of corresponding machine learning tasks and discards information that is useless or even harmful to task performance. Sparse learning is a powerful tool to achieve information selection. In this thesis, we apply sparse learning to two major areas in machine learning -- feature selection and transfer learning.

Feature selection is a dimensionality reduction technique that selects a subset of representative features. Recently, feature selection combined with sparse learning has attracted significant attention due to its outstanding performance compared with traditional feature selection methods that ignore correlation between features. However, they are restricted by design to linear data transformations, a potential drawback given that the underlying correlation structures of data are often non-linear. To leverage more sophisticated embedding than the linear model assumed by sparse learning, we propose an autoencoder-based unsupervised feature selection approach that leverages a single-layer autoencoder for a joint framework of feature selection and manifold learning. Additionally, we include spectral graph analysis on the projected data into the learning process to achieve local data geometry preservation from the original data space to the low-dimensional feature space.

Transfer learning describes a set of methods that aim at transferring knowledge from related domains to alleviate the problems caused by limited/no labeled training data in machine learnig tasks. Many transfer learning techniques have been proposed to deal with different application scenarios. However, due to the differences in data distribution, feature space, label space, etc., between source domain and target domain, it is necessary to select and only transfer relevant information from source domain to improve the performance of target learner. Otherwise, the target learner can be negatively impacted by the weak-related knowledge from source domain, which is referred to as negative transfer. In this thesis, we focus on two transfer learning scenarios for which limited labeled training data are available in target domain. In the first scenario, no label information is avaible in source data. In the second scenario, large amounts of labeled source data are available, but there is no overlap between the source and target label spaces. The corresponding transfer learning technique to the former case is called \emph{self-taught learning}, while that for the latter case is called \emph{few-shot learning}. We apply self-taught learning to visual, textal, and audio data. We also apply few-shot learning to wearable sensor based human activity data. For both cases, we propose a metric for the relevance between a target sample/class and a source sample/class, and then extract information from the related samples/classes for knowledge transfer to perform information selection so that negative transfer caused by weakly related source information can be alleviated. Experimental results show that transfer learning can provide better performance with information selection.

Share

COinS