This paper proposes a novel network compression method called dynamic network surgery, which reduces network complexity by on-the-fly connection pruning. Unlike previous methods that use a greedy approach, dynamic network surgery integrates connection splicing into the entire process to avoid incorrect pruning and enables continual network maintenance. The method is effective, achieving significant parameter compression without accuracy loss. For example, it compresses LeNet-5 by a factor of 108× and AlexNet by 17.7×, outperforming recent pruning methods. The method involves two key operations: pruning and splicing, which are integrated through dynamic updates of parameter importance. Pruning reduces network complexity, while splicing allows recovery of pruned connections if they are found to be important. This dynamic approach enables the network to adaptively maintain its structure, improving both compression efficiency and learning performance. The method is validated through experiments on various network models, including the XOR problem, MNIST database, ImageNet, and AlexNet, demonstrating its effectiveness in reducing model size while maintaining high accuracy. The results show that dynamic network surgery achieves superior compression rates compared to existing methods, with significant improvements in both storage requirements and computational efficiency. The method is implemented using stochastic gradient descent and is designed to be efficient and adaptable to different network structures.This paper proposes a novel network compression method called dynamic network surgery, which reduces network complexity by on-the-fly connection pruning. Unlike previous methods that use a greedy approach, dynamic network surgery integrates connection splicing into the entire process to avoid incorrect pruning and enables continual network maintenance. The method is effective, achieving significant parameter compression without accuracy loss. For example, it compresses LeNet-5 by a factor of 108× and AlexNet by 17.7×, outperforming recent pruning methods. The method involves two key operations: pruning and splicing, which are integrated through dynamic updates of parameter importance. Pruning reduces network complexity, while splicing allows recovery of pruned connections if they are found to be important. This dynamic approach enables the network to adaptively maintain its structure, improving both compression efficiency and learning performance. The method is validated through experiments on various network models, including the XOR problem, MNIST database, ImageNet, and AlexNet, demonstrating its effectiveness in reducing model size while maintaining high accuracy. The results show that dynamic network surgery achieves superior compression rates compared to existing methods, with significant improvements in both storage requirements and computational efficiency. The method is implemented using stochastic gradient descent and is designed to be efficient and adaptable to different network structures.