Adaptive Knowledge-Guided Pruning Algorithm AKGP with Dynamic Weight Allocation for Model Compression

In this paper, we propose the Adaptive Knowledge-Guided Pruning Algorithm (AKGP), a novel approach to model compression that enhances traditional pruning by incorporating a dynamic, data-driven weight allocation strategy during knowledge distillation. Unlike existing methods, such as the Geometric Median-based pruning approach combined with knowledge distillation and quantization proposed. AKGP dynamically balances the influence of teacher networks and real labels based on dataset characteristics. This adaptive strategy ensures that pruned models achieve superior accuracy even at high compression rates, while significantly reducing model size and computational complexity. Experimental results on the CIFAR-10 dataset demonstrate that AKGP achieves a model accuracy of 94% for ResNet 32 under a 50% pruning ratio, surpassing the baseline and previous methods. This improvement opens new possibilities for deploying deep learning models on resource-constrained devices such as mobile and embedded platforms.