Global Sensitivity Analysis and Low Magnitude Pruning for Convolutional Neural Networks Reduction in Image Net Based on Transfer Learning State of the Art Models

Transfer learning is a technique for reapplying to a new task, the knowledge held by a neural network. With constantly growing network architectures, pruning algorithms are being researched with the intent of compressing data while having the least impact on accuracy. This article investigates the Global Sensitivity Analysis (GSA) method for creating generalized pruned networks for transfer learning, and compares it with the Low Magnitude Pruning (LMP) technique. These techniques reduce popular convolutional neural networks: ResNet50V2, ResNet152V2, VGG16, VGG19 and InceptionV3, when pre-trained on ImageNet as applied to classify cifar10, cifar100 and fashion MNIST datasets in benchmark applications.