Neural Architecture Search Explained 

Thanks

Glad you like it.

No, it's not grid search. we are not trying each and every possible combination of layers in the model, the combination of filters inside the network is proposed by the controller. The controller tries to purpose these combinations in such a way so that reward which is validation Accuracy of the child network is highest. The RNN network which is purposing the network architecture after each iteration changes its weights according to the reward(which is the validation accuracy of child network). So, Instead of trying each and every combination which happened in a grid search, In NAS after some iterations on the controller, it will not purpose sub-optimal architecture