Towards a configurable and non-hierarchical search space for NAS

Neural Architecture Search (NAS) outperforms handcrafted Neural Network (NN) design. However, current NAS methods generally use hard-coded search spaces, and predefined hierarchical architectures. As a consequence, adapting them to a new problem can be cumbersome, and it is hard to know which of the NAS algorithm or the predefined hierarchical structure impacts performance the most. To improve flexibility, and be less reliant on expert knowledge, this paper proposes a NAS methodology in which the search space is easily customizable, and allows for full network search. NAS is performed with Gaussian Process (GP)-based Bayesian Optimization (BO) in a continuous architecture embedding space. This embedding is built upon a Wasserstein Autoencoder, regularized by both a Maximum Mean Discrepancy (MMD) penalization and a Fully Input Convex Neural Network (FICNN) latent predictor, trained to infer the parameter count of architectures. This paper first assesses the embedding's suitability for optimization by solving 2 computationally inexpensive problems: minimizing the number of parameters, and maximizing a zero-shot accuracy proxy. Then, two variants of complexity-aware NAS are performed on CIFAR-10 and STL-10, based on two different search spaces, providing competitive NN architectures with limited model sizes.