I recently read the "paper NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search", which can be found here.

I can say that I understood most of the paper but I am not sure I was able to grasp the main motivational idea behind the paper.

I understand that the authors choose a cell configuration and benchmarked that configuration for 15,625 candidates, keeping detailed logs for each of them. To that end I understand that the authors made it extremely easy to query the scores of different configurations and get the respective logs.

As I understand it NAS is quite expensive in terms of computation so one practitioner could not easily run something like that on a normal laptop. This leads me to believe that now one can easily get some cell-configurations that performed well on the datasets the authors tested and use them on their own networks without having to do the search themselves. Is this the motivation behind the paper or am I missing something here?

Finally, it is mentioned that the paper enables researchers to avoid unnecessary repetitive training for selected candidate and focus solely on the search algorithm itself. Does this mean that the paper enables researchers to build a search algorithm that finds the best cell configuration in the 15,625 candidates and then extend that algorithm to other cell-spaces?

I'm quite sorry if the points I'm making here sound somewhat confusing; I confess that I'm a bit inexperienced in NAS.



You must log in to answer this question.

Browse other questions tagged .