Original Paper
Introduction
Here are some questions we want to ask:
- We know certain things already affect model performance: convolution, size, depth, residuals
- This gives us a space of possible design choices to work with
- Well-optimized networks can be challenging!
NAS (Neural Architecture Search) shows good results!
(34, 18, 29)
Limitations:
- NAS produces a single network tuned to a specific setting