DARTS-PruneNet: A Two-Stage Neural Network Pruning and Architecture Reconstruction Strategy

Deep Neural Networks (DNNs) excel in computer vision but struggle on resource-constrained edge devices due to high computational and memory demands. Network pruning reduces model size by removing redundant weights but can lead to significant accuracy loss when pruning 80-99% of weights. Traditional DNN architectures are suboptimal for extreme pruning. We propose a DARTS-based approach that integrates network pruning into the architecture search process. Our method introduces PrunedConv and PrunedLinear operations to simulate pruning during the search and modifies the NAS objective to account for sparsity. The result is a pruning-aware architecture search framework that jointly optimizes network weights, pruning masks, and architecture parameters. Our approach outperforms state-of-the-art pruning methods on CIFAR-10 and ImageNet, achieving high accuracy with significantly fewer parameters. For instance, ResNet18 achieves 70.55% Top-1 accuracy on ImageNet after pruning. These results demonstrate that joint backbone reconstruction and pruning can greatly enhance the efficiency of deep networks for edge devices.