Layer-wise adaptive structured pruning via genetic algorithms with taylor-based proxy fitness

Abstract

Deploying deep convolutional neural networks (CNNs) on edge devices requires balancing model accuracy and computational efficiency. While structured pruning limits inference costs by removing redundant filters, most methods apply a rigid, global criterion, ignoring the distinct representational roles of individual layers. This yields suboptimal results, especially under aggressive compression where over-pruning degrades performance. To address this limitation, we propose an adaptive structured pruning framework based on genetic algorithms (GAs) that jointly optimizes layer-wise pruning ratios and strategies. Each layer independently selects between min-importance and median-rank pruning, enabling the exploration of tailored strategy combinations. A training-free taylor based proxy fitness function ensures efficient candidate evaluation without re peated fine-tuning. After fine-tuning the selected architecture, experiments on VGG16 demonstrate that our method achieves 92.78 ± 0.28% accuracy (over 50 independent runs) with a 70.0 ± 3.2% MACsreductiononCIFAR-10, andmaintains 71.82% accuracy on CIFAR-100. These results demonstrate competitive performance compared to existing pruning methods while achieving substantial computational cost reduction.