Forget the name for a second. Building a neural network normally means a human deciding how many layers it has, how wide each one is, and how they connect. Neural architecture search automates that decision. You define a space of possible architectures, then run a search procedure that proposes a design, trains it a little, scores it, and uses the score to propose a better one next time.
The 2020 publication US20200334416A1 sits in this lineage. Its method couples a representation-learning model to a downstream task and iterates over configurations to improve performance on the target objective. The CPC tags include G06N 3/08, the class for neural-network learning methods, which is exactly where you'd expect a NAS-flavored method to land.
The way this actually works is a feedback loop with two clocks. The inner clock trains a candidate network. The outer clock evaluates that candidate and adjusts the search. Early NAS systems were brutally expensive because every candidate needed real training; later refinements like weight sharing and proxy evaluations cut the cost, which is why the technique went from research curiosity to something you'd see in a production patent by 2020.
Why does a general reader care? Because NAS is one of the quiet reasons the models you use keep getting better without a corresponding army of architects. The design work is increasingly done by search, not by hand. That shifts the bottleneck from human intuition to compute, a theme that runs through the entire AI buildout story.
A caveat in the spirit of the house: a publication is a published application, not a granted patent, and it describes a method, not a benchmark you can trust blindly. What it does establish is that by 2020, automating model design was mature enough to write down and claim. Strip the adjectives and that's the signal worth keeping.