RQ1

Experiment Setup 

In RQ1, we compare the following three slicing methods: global slicing, category-wise slicing, and semantic slicing. To assess neuron contribution and select critical neurons, we utilize five neuron contribution metrics: Avg, Var, Ens, DeepLIFT, and Taylor. As outlined in Section III-A of paper, global slicing and category-wise slicing choose top 10%, 20%, ... up to 100% of neurons in each layer or category for slicing respectively. Semantic slicing identifies critical neurons for each category using a set of thresholds Θ = 0.9, 0.91, 0.92, 0.93, 0.94, 0.95, 0.96, 0.97, 0.98, 0.99. Then it merges them into semantic slices, as described in Section III-C. 

For RQ1, the compression rate is defined as the percentage reduction of masked parameters in the model, while model accuracy denotes the predictive accuracy of the model slices on the test dataset. We conducted comparisons of slicing precision across the previously mentioned six experimental settings.

MLP-MNIST

MLP-FMNIST

AlexNet-CIFAR10

VGG16-CIFAR10

ResNet101-CIFAR10

VGG19-ImageNet