This post summarize the work Differentiable Architecture Search (Liu et al, 2018) - CNN Part.
As the name suggest, alongside with other AutoML/NAS(Neural Architecture Search) approachs using Reinforcement Learning, Bayes Optimization or Evolutionary algorithm, DARTS uses gradient-based method by relaxing the search space to continuous, then jointly optimize architecture parameters $\alpha$ and weight parameters $\mathbf{w}$ (bi-level optimization).
The authors observed that the typical Convolutional Neural Network is the repetition of some motifs/blocks/cells (like Resisual block in ResNet, Inception Module in InceptionNet). So, instead of find the best network, they only find the optimial cell.
In DARTS, a cell is a Directed Acyclic Graph (DAG), it includes $k$ nodes, each node connects to preceding nodes. Each edge that connects a pair of nodes $(i, j)$, applies some predefined operations (convolution, seperable convolution, pooling, and more).
After training finish, the authors derive the final architecture from architecture parameters $\alpha$.
