摘要： 
Topic: 

Algorithm discovery as program search.
The discovered algorithm used for Neural Network training.

Problem: 

infinite and sparse program space → efficient search techniques
gap between proxy and target space → selection and simplification strategy

Advantages: 

memory efficient
better performance compared with Adam (applied to many model training tasks)

Analysis (better performanc...