แนวคิดหลัก
DRIVE leverages a novel dual gradient-based metric to rapidly prune deep neural networks while maintaining high accuracy, bridging the gap between exhaustive and initialization-based pruning methods.
บทคัดย่อ
The paper introduces Dual Gradient-Based Rapid Iterative Pruning (DRIVE), a novel early pruning technique for deep neural networks (DNNs). DRIVE aims to address the trade-off between accuracy and pruning time observed in existing pruning methods.
The key highlights are:
DRIVE starts by training the unpruned model for a few epochs, allowing essential parameters to acquire larger magnitudes and indicating their significance.
DRIVE's pruning metric combines three key components:
Parameter magnitude (L1 norm)
Connection sensitivity, which captures the impact on the loss when a parameter is removed
Convergence sensitivity, which considers the proximity of the parameter to convergence
The dual gradient-based metric in DRIVE helps identify and preserve parameters that may not be currently important but could become essential as training progresses, addressing the limitations of initialization-based pruning methods.
Experiments on various DNN architectures (AlexNet, VGG-16, ResNet-18) and datasets (CIFAR-10, CIFAR-100, Tiny ImageNet, ImageNet) show that DRIVE consistently outperforms the accuracy of initialization-based pruning methods (SNIP, SynFlow) while being 43x to 869x faster than the computationally intensive iterative magnitude pruning (IMP) method.
DRIVE bridges the gap between the speed of pruning and the accuracy of the sparse networks produced by exhaustive pruning, offering a viable solution to address the energy challenge in training large-scale models by leveraging sparsity from the onset.
สถิติ
The paper presents several key metrics and figures to support the authors' claims:
"DRIVE is 43× to 869× faster than IMP for pruning."
"DRIVE consistently surpasses the accuracy of initialization-based pruning methods (SNIP, SynFlow) across various networks and datasets."
คำพูด
"DRIVE leverages a novel dual gradient-based metric to rapidly prune deep neural networks while maintaining high accuracy, bridging the gap between exhaustive and initialization-based pruning methods."
"DRIVE bridges the gap between the speed of pruning and the accuracy of the sparse networks produced by exhaustive pruning, offering a viable solution to address the energy challenge in training large-scale models by leveraging sparsity from the onset."