Pruning Followed the Model's Geometry

Setting a neural-network parameter to zero moves the model onto a constrained surface. This paper measures that displacement with the Fisher information geometry, then derives a hierarchy from simple magnitude pruning toward progressively more faithful geodesic approximations. Across fully connected networks and vision transformers on MNIST and CIFAR-10, over the complete zero-to-100% pruning range and five seeds, the proposed schemes outperformed magnitude and local-Fisher pruning on both accuracy and Matthews correlation. Intermediate approximations retained much of the gain with lower computational cost; the evidence remains limited to the reported models and datasets.