Relative Repairability: A Calibration-Based Diagnostic for High-Sparsity Post-Pruning Allocation
Researchers have developed a new diagnostic tool called Relative Repairability (RR) to help optimize neural network pruning, particularly at high sparsity levels. RR assesses how much damage from pruning can be recovered by a lightweight repair procedure, using unlabeled calibration data. Experiments on ResNet and VGG models showed RR is most effective near a specific transition point where standard pruning methods become less reliable, suggesting that high-sparsity pruning should consider both weight retention and repairability. AI
IMPACT Introduces a novel method to improve the efficiency and effectiveness of neural network pruning, potentially leading to smaller, faster models.