Table 1 Comparison of the architectures with the lowest test error (averaged over 200 trials) returned by NAS algorithms after 150 architecture evaluations on NASBench-101. \(^{\dagger }\): results from BANANAS.
NAS algorithm | Method | Test error |
|---|---|---|
REINFORCE\(^{\dagger }\) | Reinforcement learning | 6.436 |
TPE\(^{\dagger }\) | BO (Parzen windows) | 6.415 |
BOHB\(^{\dagger }\) | BO (successive halving) | 6.356 |
Random search\(^{\dagger }\) | Random search | 6.341 |
GCN Pred.\(^{\dagger }\) | GCN | 6.331 |
BO w. GP\(^{\dagger }\) | BO (Gaussian process) | 6.267 |
NASBOT\(^{\dagger }\) | BO (Gaussian process) | 6.250 |
AlphaX\(^{\dagger }\) | Monte Carlo tree search | 6.233 |
Reg. Evolution\(^{\dagger }\) | Evolution | 6.109 |
DNGO\(^{\dagger }\) | BO (neural networks) | 6.085 |
BOHAMIANN\(^{\dagger }\) | BO (Bayesian NN) | 6.010 |
BONAS\(^{\dagger }\) | BO (GCN) | 5.954 |
Local search\(^{\dagger }\) | Local search | 5.932 |
BANANAS\(^{\dagger }\) | BO (path encoding) | 5.923 |
AE-NAS(Ours) | Attention mechanism | 5.923 |
