“DARTS: Differentiable Architecture Search” was submitted by Hanxiao Liu, Karen Simonyan, and Yiming Yang in June 2018. It attacked the central problem of early neural architecture search, which is that searching over a discrete space of architectures with reinforcement learning or evolution was extremely slow, often costing thousands of GPU-days.
DARTS reformulates the search so it becomes differentiable. Instead of picking one operation per connection, it places a weighted mixture of all candidate operations on each edge of the network and learns the mixing weights by ordinary gradient descent, jointly with the network’s own parameters. After search, the highest-weighted operation on each edge is kept, yielding a concrete architecture. This continuous relaxation let DARTS find competitive architectures on CIFAR-10, ImageNet, Penn Treebank, and WikiText-2 in roughly a day on a single GPU, orders of magnitude faster than the reinforcement-learning approach.
DARTS became one of the most influential and widely extended NAS methods, and the authors released their code publicly. Later work noted stability and reproducibility issues, prompting many DARTS variants.
For a business reader, DARTS represents the step that turned architecture search from a research-lab luxury into something a small team could realistically run.