一种基于权重重构的忆阻神经网络剪枝方法OA

Resistive Random Access Memory(RRAM),with its inherent processing-in-memory capability,has emerged as an efficient hardware platform for neural network acceleration.Pruning techniques effectively compress neural networks by removing redundant weights,thereby reducing the hardware cost of RRAM-based accelerators.However,existing RRAM-oriented structured pruning methods often suffer from exces-sively coarse granularity,which can lead to accuracy degradation.Moreover,they typically neglect the nu-merical patterns shared among weights,leaving potential redundancy underexploited and limiting further im-provements in compression and hardware efficiency.To address these challenges,we propose a pruning method for RRAM neural networks based on Weight Reconstruction.Specifically,an integer scaling weight reconstruction strategy is designed to extract and share common numerical structures among weights,while discarding components with minimal impact on model accuracy.The essential weight information is then mapped onto the RRAM crossbar for network inference,achieving a compact weight representation.Further-more,a progressive retraining mechanism is introduced,where the discarded components are leveraged as guidance signals that are gradually attenuated to refine the model,thereby recovering accuracy while maintain-ing high compression and hardware efficiency.Experiments show that,compared to the state-of-the-art works,the proposed method achieves up to 1.2×,1.2×,and 1.3×improvements in compression rate,area efficiency,and energy efficiency,respectively,with negligible accuracy loss.