Abstract: Deep convolutional neural networks (CNNs) have demonstrated its extraordinary power on various visual tasks like object detection and classification. However, it is still challenging to deploy state-of-the-art models into real-world applications, such as autonomous vehicles, due to their expensive computation costs. In this paper, to accelerate the network inference, we introduce a novel pruning method named Drop-path to reduce model parameters of 2D deep CNNs. Given a trained deep CNN, pruning paths with different lengths is achieved by ordering the influence of neurons in each layer on the probably approximately correct (PAC) Bayesian boundary of the model. We believe that the invariance of PAC-Bayesian boundary is an important factor to guarantee the generalization ability of deep CNN under the condition of optimizing as much as possible. To the best of our knowledge, this is the first time to reduce model size based on the generalization error boundary. After pruning, we observe that the convolutional kernels themselves become sparse, rather than some being removed directly. In fact, Drop-path is generic and can be well generalized on multi-layer and multi-branch models, since parameter ranking criterion can be applied to any kind of layer and the importance scores can still be propagated. Finally, Drop-path is evaluated on two image classification benchmark datasets (ImageNet and CIFAR-10) with multiple deep CNN models, including AlexNet, VGG-16, GoogLeNet, and ResNet-34/50/56/110. Experimental results demonstrate that Drop-path achieves significant model compression and acceleration with negligible accuracy loss.