A neural network architecture optimizer based on DARTS and generative adversarial learning

Neural network architecture search automatically configures a set of architectures according to the targeted rules. Thus, it relieves human-dependent effort and repetitive resources consumption for designing neural makes task finding optimum with better performance much more accessible. Network methods based on differentiable (DARTS), however, introduces parameter redundancy. To address this issue, work presents novel method optimizing that combines DARTS generative adversarial learning (GAL). We first find module structures utilizing algorithm. Afterwards, retrieved modules are stacked derive initial architecture. Next, GAL is used prune some branches network, thereby obtaining final The proposed DARTS-GAL re-optimizes searched by simplify connection reduce parameters without compromising performance. Experimental results benchmark datasets, i.e., Mixed National Institute Standards Technology (MNIST), FashionMNIST, Canadian Advanced Research10 (CIFAR10), Research100 (CIAFR100), Cats vs Dogs, voiceprint recognition indicate test accuracies higher than those in majority cases. In particular, solution exhibits an improvement accuracy 7.35% CIFAR10 compared DARTS, attaining state-of-the-art result 99.60%. Additionally, number derived significantly lower method, pruning rate 62.3% at highest case.