Fisher-Rao guided channel pruning with progressive re-estimation

Structured channel pruning enables dense, deployment-friendly model compression, but its reliability depends on the channel-importance criterion and pruning schedule. We propose a Fisher-Rao guided framework that scores channel gates with a diagonal empirical Fisher approximation and combines this score with progressive re-estimation and short recovery training. On CIFAR-10, it reduces parameters/FLOPs by 71.0%/56.2% on ResNet-56, 85.5%/66.8% on ResNet-110, 90.2%/50.6% on VGG-16, and 74.8%/57.7% on GoogLeNet. On ImageNet, it reaches 75.65 ± 0.04% Top-1 with 60.0%/50.1% reduction on ResNet-50, 71.59 ± 0.05% with 60.0%/54.4% reduction on MobileNetV2, and 75.90 ± 0.07% with 77.2%/50.4% reduction on EfficientNet-B0. Recovery-free diagnostics show backbone-dependent sensitivity, positioning the method as a loss-coupled and budget-controlled classification pruning baseline rather than a hardware or task-specific system.