Fisher Information Metric Optimizes Neural Architecture Pruning

A new pruning scheme uses the geodesic distance in model space, derived from the Fisher information metric, to determine parameter importance. The method demonstrates higher optimality than traditional magnitude pruning for fully-connected networks and vision transformers.