A Systematic Overview of Meta-pruning Strategies in Deep Learning

This survey paper examines meta-pruning techniques in deep learning, aimed at optimizing deep neural networks (DNNs) for efficient deployment on edge devices. DNNs play a vital role in computer vision and AI but face challenges due to their computational complexity and large model size. Meta-pruning leverages meta-learning principles to intelligently select and fine-tune pruned architectures, ensuring compactness while maintaining high accuracy. This approach differs from traditional pruning methods and can address issues such as performance degradation and sensitivity to initializations. The paper starts by discussing the drawbacks of deploying DNNs on resource-constrained edge devices and then explores model pruning as a solution. It delves into the complexities of meta-learning and how it might be applied in the context of meta-pruning. The paper concludes with empirical experiments demonstrating the effectiveness of various meta-pruning strategies. Overall, this survey provides a comprehensive understanding of how meta-pruning enhances the efficiency and performance of DNNs, especially in edge computing scenarios.