Neural Architecture Search
Automatically searching a defined space of possible network architectures, using a search strategy plus a cost-aware score, instead of hand-designing one.
A search space defines which layer choices are candidates, a search strategy proposes architectures from that space, and a cost model scores each one. Naive search trains every candidate from scratch, which becomes infeasible as the space grows exponentially with layer count; weight sharing trains one supernet containing every candidate operation instead, decoupling search cost from search space size.
The final selected architecture should still be retrained from scratch and verified, since weight-sharing performance estimates can rank candidates differently than a standalone training run would.