Abstract

With the ever-increasing demand for the quality and quantity of the training samples, it is difficult to replicate the success of modern machine learning models in knowledge-rich domains, where the labeled data for training is scarce and labeling new data is expensive. While machine learning and AI have achieved significant progress in many common domains, the lack of large-scale labeled data samples poses a grand challenge for the wide application of advanced statistical learning models in key knowledge-rich domains, such as medicine, biology, physical science, and more. Active learning (AL) offers a promising and powerful learning paradigm that can significantly reduce the data-annotation stress by allowing the model to only sample the informative objects to learn from human experts. Previous AL models leverage simple criteria to explore the data space and achieve fast convergence of AL. However, those active sampling methods are less effective in exploring knowledge-rich data spaces and result in slow convergence of AL. In this thesis, we propose novel AL methods to address knowledge-rich data exploration challenges with respect to different types of machine learning tasks. Specifically, for multi-class tasks, we propose three approaches that leverage different types of sparse kernel machines to better capture the data covariance and use them to guide effective data exploration in a complex feature space. For multi-label tasks, it is essential to capture label correlations, and we model them in three different approaches to guide effective data exploration in a large and correlated label space. For data exploration in a very high-dimension feature space, we present novel uncertainty measures to better control the exploration behavior of deep learning models and leverage a uniquely designed regularizer to achieve effective exploration in high-dimension space. Our proposed models not only exhibit a good behavior of exploration for different types of knowledge-rich data but also manage to achieve an optimal exploration-exploitation balance with strong theoretical underpinnings. In the end, we study active learning in a more realistic scenario where human annotators provide noisy labels. We propose a re-sampling paradigm that leverages the machine's awareness to reduce the noise rate. We theoretically prove the effectiveness of the re-sampling paradigm and design a novel spatial-temporal active re-sampling function by leveraging the critical spatial and temporal properties of the maximum-margin kernel classifiers.

Publication Date

7-2022

Document Type

Dissertation

Student Type

Graduate

Degree Name

Computing and Information Sciences (Ph.D.)

Advisor

Qi Yu

Advisor/Committee Member

Yu Kong

Advisor/Committee Member

Ifeoma Nwogu

Campus

RIT – Main Campus

Share

COinS