DARTS-EAST: an edge-adaptive selection with topology first differentiable architecture selection method

Xuwei Fang, Weisheng Xie, Hui Li, Wenbin Zhou, Chen Hang, Xiangxiang Gao

Abstract: DARTS+PT is a well-known differentiable neural architecture search (NAS) method that evaluates the contribution of operations to the performance of the super-network, ultimately deriving the final architecture. However, DARTS+PT introduces randomness into the edge discretization process by selecting edges randomly, which leads to performance instability. Moreover, the method assesses the impact of each candidate operation by iteratively removing them and measuring the resulting drop in super-network performance, leading to a high search cost. To address these issues, this paper identifies the root cause of instability and proposes a novel edge selection criterion to establish an adaptive edge discretization order, improving stability. Additionally, we introduce a topology-first discretization scheme that prioritizes topology selection over operation selection, significantly reducing the search cost. We name this approach DARTS-EAST (Edge-Adaptive Selection with Topology-First Differentiable Architecture Selection). Extensive experiments on widely used benchmarks demonstrate that DARTS-EAST not only achieves competitive performance but also offers significant improvements in both stability and search efficiency.