Definition
Plain language
Methods that search for the best neural network design automatically, rather than having humans hand-tune it.
As stated in the literature
Neural Architecture Search — a family of automated methods for discovering high-performing network architectures via reinforcement learning, evolutionary search, gradient-based methods, or, more recently, LLM-driven agentic search.
Also called: Neural Architecture Search
Why it matters: Hand-tuning architectures is slow and biased by tradition, while automated search can find designs humans wouldn't have tried.
For example, a NAS system might propose, train, and rank thousands of small vision networks to find one that beats a hand-designed baseline.
Heard on the show
“Traditional Neural Architecture Search uses Bayesian optimization or evolutionary algorithms — but those are rigid, mechanical procedures.”Episode 053 — An AI Agent Swapped In Focal Loss And Beat A Human-Tuned Training Script