Neural Architecture Search

Neural architecture search (NAS) is the automated design of a neural network’s structure, the number and type of layers, how they connect, and the operations inside them, by treating that design as a search or optimization problem rather than a task for human intuition. It is the deep-learning-specific branch of AutoML.

The concept was popularized by the 2016 paper “Neural Architecture Search with Reinforcement Learning” by Zoph and Le, which trained a controller network to propose architectures and rewarded it for the accuracy of the resulting models. Because evaluating each candidate meant training a network, early NAS was extremely expensive, which drove a wave of efficiency research: weight-sharing approaches that reuse computation across candidates, and differentiable methods such as DARTS that relax the discrete search into a continuous one solvable by gradient descent. Reproducibility benchmarks like NAS-Bench-101 later let researchers compare methods fairly and cheaply. NAS has produced competitive architectures in vision and language and fed into products such as Google’s image models.

For a business reader, NAS represents the field automating one of its most specialized crafts, with the practical caveat that the search itself can consume significant compute.

Sources

Last verified June 7, 2026