Towards understanding why Lookahead generalizes better than SGD and beyond

• Main Authors: ZHOU, Pan, YAN, Hanshu, YUAN, Xiaotong, FENG, Jiashi, YAN, Shuicheng
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2021
• Subjects: Theory and Algorithms
• Online Access: https://ink.library.smu.edu.sg/sis_research/8987; https://ink.library.smu.edu.sg/context/sis_research/article/9990/viewcontent/2021_NeurIPS_lookahead.pdf
• Institution: Singapore Management University

To train networks, lookahead algorithm [1] updates its fast weights k times via an inner-loop optimizer before updating its slow weights once by using the latest fast weights. Any optimizer, e.g. SGD, can serve as the inner-loop optimizer, and the derived lookahead generally enjoys remarkable test p...