Fast Post-Training Pruning Framework for CNNs Based on Fisher Information

This paper presents a novel post-training pruning framework for Convolutional Neural Networks (CNNs) that achieves superior computational efficiency while maintaining model accuracy. My approach introduces three key innovations: (1) a Fisher information-based importance metric specifically adapted for CNNs [1], (2) a greedy mask rearrangement algorithm that accounts for inter-filter dependencies through block-diagonal Hessian approximations, and (3) a lightweight activation reconstruction method using learned scaling factors [2]. Through extensive experiments on MNIST with varying network architectures, I demonstrate that my framework achieves up to 37.29% FLOPs reduction with only 2.73% accuracy drop, while completing the entire pruning process in under 3 minutes on a single GPU. Comparative analysis shows my method reduces accuracy drop by 28-55% compared to magnitude-based pruning (threshold=0.1) and L1 regularization (λ=0.001) approaches]. The mathematical analysis provides theoretical guarantees on error bounds, while practical guidelines address real-world deployment considerations.