Differentiable Architecture Search Algorithm Based on Global Comparison

Manually building neural network models is a great test of researchers’ knowledge reserves, so using Neural Architecture Search (NAS) to automatically construct networks becoming increasingly popular. This paper uses an improved Differentiable (DARTS) build classification model, DARTS gradient-based NAS algorithm. However, local selection, this proposes selection algorithm based on global selection. The can ensure that all operations connected the same intermediate node be fairly compared, ensuring fairness in and searching for more types architectures. only fixed candidate network, structure relatively single, skip connection will dominate searched large number approximation strategies, which easily lead decrease accuracy model. For problems, we add SENet attention mechanism after cell output, extract features feature layer channel dimension, it not improve search performance but also effectively increase diversity robustness network. final error CIFAR10 reaches 2.48%.