Lightweight Generative AI on Edge Devices: Pruning Strategies for VGG-16 and MobileNet on CIFAR

Emerging deep learning advancements have fueled the desire to run models on edge devices, where both memory and computational power are very limited. This research explores pruning methods for compressing deep neural networks to deploy them at the edge, emphasizing filter pruning and initialization pruning. The work compares the effect of these methods on VGG-16 and MobileNet models with CIFAR-10 and CIFAR-100 datasets. Findings reveal that filter pruning is capable of reducing the model size by up to 80% with small accuracy loss, while pruning at initialization retains accuracy better in models of small size. The observations suggest the ability of pruning mechanisms to optimize efficiency in edge AI.