RQ2

Experiment Setup

To systematically assess NeuSemSlice for the model restructuring task, we adopt the same experimental setup detailed in DeepArc, and validate by comparing the compression efficiency and effects on the model before and after applying NeuSemSlice. Specifically, after removing the non-critical neurons of the model, we expect the compression technology to directly focus on the redundant parts within the critical neurons, improving the iteration efficiency. Based on this, we apply four of the most advanced model compression techniques currently available—LAMP, Global, Uniform+, and ERK—for optimal compression results.

In our experiments, we take the following metrics to evaluate the performance of model compression. The Compression Rate (CR) indicates the reduction in parameter size, the Number of Iterations (NI) reflects the efficiency of the compression process, and Model Accuracy (MA) measures the predictive performance on the test dataset.

MLP-MNIST

MLP-FMNIST

AlexNet-CIFAR10

VGG16-CIFAR10

ResNet101-CIFAR10

VGG19-ImageNet