Arash Jamshidi

This thesis focuses on the use of greedy methods in machine learning and data analysis. In this research, we plan to utilize relevant information hidden in the structure of learning problems, such as gradients in continuous spaces and the combinatorial properties of discrete learning problems, to design faster and more reliable algorithms with a minimal amount of data. We aim to design algorithms that employ greedy procedures that (i) require fewer samples than standard methods or (ii) consider only a small fraction of the search space. Possible projects include early stopping in gradient descent, designing faster differentially private mechanisms, and data-efficient machine learning.