Shapley value optimized differentiable architecture search for lightweight neural networks in resource-constrained environments.

In resource-constrained environments such as embedded systems, IoT devices, and underwater equipment, efficient neural networks with low computational overhead are essential. Differentiable Architecture Search (DARTS) enables architecture optimization but struggles to reflect the importance of various operations in a large search space. This paper proposes a Shapley value-based method for optimizing the search space in DARTS to fairly measure the contribution of each operation type in architecture search. By integrating operation types with their corresponding connection nodes, the method reduces unnecessary operations and compresses the search space, leading to lightweight networks and enhanced efficiency.Experiments have shown that on the CIFAR-10 dataset, this method only requires 2.7 M parameters and a search cost of 0.26GPU days, achieving a competitive testing accuracy of 95.45 %. On the ImageNet dataset, this method only requires 4.3 M parameters and has a search cost of 12.2 GPU days, achieving an accuracy of 76.14 %. The experimental results are superior to other comparative algorithms. This method significantly improves the efficiency of architecture search while maintaining low topology complexity, and outperforms existing NAS methods.