What is Neural Architecture Search (NAS) weight sharing and why is it important?

Weight sharing allows multiple candidate architectures in NAS to reuse parameters, drastically reducing training cost.

Traditional NAS evaluates each architecture independently, which is extremely expensive. Weight-sharing NAS trains a supernet containing all candidate operations. Each sub-network shares weights from this supernet, enabling rapid evaluation without full retraining. Methods like ENAS and DARTS use this approach to make architecture search computationally feasible.