Work

More Robust Learning Guarantees via Active Dataset Construction

Year: 2025
Type: dissertation
Abstract: The widespread adoption of machine learning continues to give rise to settings in whichtargeted sampling and labeling can be leveraged during training. Simultaneously, the increased reliance on machine learning systems makes more robust performance guarantees highly desir- able. These factors motivate the study of a vein of modern learning problems in which more precise accounting is required to accurately represent the cost of training data, and PAC-learning guarantees stronger than those traditionally studied are the goal. This dissertation largely explores this interaction between active dataset construction and robust machine learning.We begin with a study exploring the potential of active labeling to improve accuracy guarantees when natural nonparametric methods are used for classifier construction. This servesto contextualize the connection to active learning as traditionally studied in the literature. We then consider the first of two truly “robust” machine learning problems: the problem of learning over a collection of distributions, where worst-case accuracy is the learning objective. We study this problem from the perspective of active data acquisition, assuming the learner has the power to construct a training set of labeled examples from each distribution in the collection. Finally, we turn to the problem of learning a generative model that should produce examples meeting some domain-specific quality criteria. Given the ability to actively query whether a given example meets these criteria, we investigate the importance of problem formulation in determining the extent to which active queries are required.
Language: English
Open Access status: green
OpenAlex ID: W7067883754
Cites: 65
Cited by:
Related to: 10