A Greedy Hierarchical Approach to Whole-Network Filter-Pruning in CNNs
Researchers have developed a novel two-level hierarchical approach for whole-network filter pruning in Convolutional Neural Networks (CNNs). This method efficiently reduces model size and computational requirements by pruning filters across all layers. The approach utilizes a sparse-approximation formulation and a novel closed-form error criterion for backward pruning, outperforming existing state-of-the-art methods on various benchmark networks like ResNet and VGG. AI
IMPACT Reduces model size and computational demands for CNNs, enabling deployment on resource-constrained devices.