Differentiable Network Pruning via Polarization of Probabilistic Channelwise Soft Masks

<div>Investigation of the consistency of learned soft masks across different input batches. For a given filter, values of the mask learned from different batches are compared. The experiments were conducted on VGG16, ResNet32, ResNet56, and ResNet110 using the CIFAR-10 dataset. The soft masks of randomly sampled 150 filters are analyzed across 150 batches.</div>

Computational Intelligence and Neuroscience

fig3

Figure 3

Figure 3: Differentiable Network Pruning via Polarization of Probabilistic Channelwise Soft Masks 