Classification of methods of optimization of cnn models

Abstract

Reducing the computational cost of CNNs without sacrificing accuracy is still one of the central problems in deep learning. This work analyzes four optimization families - pruning, quantization, knowledge distillation, and NAS - comparing their trade-offs in accuracy, compression, and hardware requirements. The analysis shows that quantization combined with pruning offers the most practical solution for edge deployment, while NAS achieves the best compression at high computational cost.