Differentiable Network Pruning via Polarization of Probabilistic Channelwise Soft Masks

<div>The change in the test accuracy of different methods with respect to the percent of reduced FLOPs. (a) shows the comparison results of pruning ResNet56 on CIFAR-10, and the pruning rate changes from 50% to 70%. (b) depicts the comparison results of pruning ResNet50 on ImageNet, and the pruning rate changes from 40% to 75%. PPSM is compared to the most representative existing methods.</div>

Computational Intelligence and Neuroscience

fig4

Figure 4

Figure 4: Differentiable Network Pruning via Polarization of Probabilistic Channelwise Soft Masks 