Zero-cost proxies for neural architecture search
Project Description
This project aims to construct more efficient zero-cost proxies to predict post-training performances of vision transformers.
Supervisor
YANG, Can
Quota
1
Course type
UROP1000
UROP1100
UROP2100
UROP3100
Applicant's Roles
The student will conduct research by exploring relevant papers, designing zero-cost proxies, testing proxies, develop mathematical justifications, and writing reports.
Applicant's Learning Objectives
The project aims to help the student learn the current theories of evaluating neural network architectures and build a zero-cost proxy.
Complexity of the project
Moderate