Picking Winning Tickets Before Training by Preserving Gradient Flow
This addresses the resource-intensive training process for large neural networks, offering a method to prune before training, which is incremental but improves upon existing pruning techniques.
The paper tackles the problem of reducing the computational cost of training overparameterized neural networks by pruning them at initialization, using a criterion called Gradient Signal Preservation (GraSP) that preserves gradient flow. The result shows that pruning 80% of weights in a VGG-16 network on ImageNet leads to only a 1.6% drop in top-1 accuracy, with better performance at extreme sparsity levels.
Overparameterization has been shown to benefit both the optimization and generalization of neural networks, but large networks are resource hungry at both training and test time. Network pruning can reduce test-time resource requirements, but is typically applied to trained networks and therefore cannot avoid the expensive training process. We aim to prune networks at initialization, thereby saving resources at training time as well. Specifically, we argue that efficient training requires preserving the gradient flow through the network. This leads to a simple but effective pruning criterion we term Gradient Signal Preservation (GraSP). We empirically investigate the effectiveness of the proposed method with extensive experiments on CIFAR-10, CIFAR-100, Tiny-ImageNet and ImageNet, using VGGNet and ResNet architectures. Our method can prune 80% of the weights of a VGG-16 network on ImageNet at initialization, with only a 1.6% drop in top-1 accuracy. Moreover, our method achieves significantly better performance than the baseline at extreme sparsity levels.