Computer Science Thesis Oral
— 5:00pm
Location:
In Person
-
Gates Hillman 8102
Speaker:
QUANG MINH HOANG
,
Ph.D. Candidate, Computer Science Department, Carnegie Mellon University
https://scholar.google.com/citations?user=56Mb6DY0_NUC
Practical Methods for Automated Algorithm Design in Machine Learning and Computational Biology
Configuration tuning is an essential practice to achieve good performance with many computational methods. However, configuring complex and discrete algorithms often requires significant trial-and-error effort due to a lack of automated solutions. In large-scale systems where computational tasks are numerous and constantly changing in specificity, the repetitive cost of manual tuning becomes a major bottleneck that hinders scalability. Moreover, the absence of a systematic approach to configure deployment settings makes it challenging to replicate the obtained results in different deploying conditions.
To address these problems, this thesis focuses on developing new data-driven automated algorithm design (AAD) frameworks in several classical and multi-task settings. Specifically, in the classical configuration tuning setting, this thesis addresses the problems of kernel selection for Bayesian methods, and minimizer construction for biological sequence sketching. In the multi-task scenario, we tackle the problems of privacy-preserving neural architecture search for multiple clients, and meta-learning for parameter optimization in a heterogeneous task stream. In all of these problems, the variables of interest often have underlying discrete structures such as trees, graphs or permutations, which result in challenging optimization problems.
The contribution of this thesis is a suite of reformulation techniques for said optimization problems. These techniques result in efficient and accurate tuning methods for configuration problems in machine learning and computational biology domains. We demonstrate the performance of our methods on various practical scenarios, and show that they have significantly outperformed the state-of-the-art benchmarks in terms of the respective design goals.
Thesis Committee:
Carl Kingsford (Chair)
David P. Woodruff
Maria-Florina Balcan
Risto Mikkulainen (University of Texas at Austin)