[RegNet] Designing Network Design Spaces (March 2020 CVPR)

[Andrew Jong]

0. Article Information and Links

• Paper link: https://arxiv.org/abs/2003.13678
• Release date: 2020/03/30
• Number of citations (as of 2020/04/20): 0
• Implementation code: PyCls, general repo, Unofficial signatrix/regnet
• Supplemental links (e.g. results):
• Publication: CVPR 2020

1. What do the authors try to accomplish?

• Develop a result-driven, human-in-the-loop process that reveals best practices for neural network architectures. This process is general, and can be applied to experimentally discover rules of thumb for (theoretically) any problem.
• This process is called Design Space Design

2. What's great compared to previous research?

• Contribution of an original process to understand network design.

Results

• more accurate + 5 times faster than EfficientNet family

3. Where are the key elements of the technology and method?

4. How did you verify that it works?

5. Things to discuss? (e.g. weaknesses, potential for future work, relation to other work)

6. Are there any papers to read next?

7. References