Adaptive Signal Resuscitation: Channel-Wise Post-Pruning Repair Prevents Accuracy Collapse in Sparse Networks

arXiv·medium signal

Paper tackles the accuracy collapse that occurs with one-shot magnitude pruning at high sparsity, even when the largest weights are preserved, by introducing a channel-wise repair technique that restores signal flow after pruning. For teams deploying pruned models to edge devices or optimizing inference costs, this provides a practical post-processing step that recovers accuracy without retraining.