Local Search is a Remarkably Strong Baseline for Neural Architecture Search

Neural Architecture Search (NAS), i.e., the automation of neural network design, has gained much popularity in recent years with increasingly complex search algorithms being proposed. Yet, solid comparisons simple baselines are often missing. At same time, retrospective studies have found many new to be no better than random (RS). In this work we consider use a Local (LS) algorithm for NAS. We particularly multi-objective NAS formulation, accuracy and complexity as two objectives, understanding trade-off between these objectives is arguably among most interesting aspects The proposed LS compared RS evolutionary (EAs), heralded ideal optimization. To promote reproducibility, create release benchmark datasets, named MacroNAS-C10 -C100, containing 200K saved evaluations established image classification tasks, CIFAR-10 CIFAR-100. Our benchmarks designed complementary existing benchmarks, especially that they suited search. additionally version problem larger architecture space. While find show considered explore space fundamentally different ways, also substantially outperforms even performs nearly good state-of-the-art EAs. believe provides strong evidence truly competitive baseline against which should benchmarked.