The current research results are as follows:
There’re five type of methods used in architecture search.
- reinforcement learning (RL),
- evolutionary algorithm (EA),
- gradient optimization (GO),
- random search (RS)
- sequential model-based optimization (SMBO).
I don’t understand much about NAS. But from above result, Neural Architectures created by machine have bigger number of parameters and lower precision. So using human-designed neural network architecture is still better for now. Hope there will be some big improvement in NAS in the near future.
There’re some other useful link