Neural Architecture Search (NAS) is an automated process in which suitable network architectures for neural networks are found using optimization algorithms.
Typical methods include reinforcement learning, evolutionary algorithms, or differentiable architecture search methods.
The search space defines possible architectures, e.g., number and type of layers, types of connections, or hyperparameters.
NAS optimizes both performance (e.g., accuracy) and efficiency (e.g., computational cost) and can thus produce better models than manually designed architectures.
An important aspect is the balance between search time and model quality, as NAS is usually very computationally intensive and techniques such as proxy models or weight sharing are used to accelerate the search.