New Fisher-Information Geodesic Pruning Beats Magnitude Pruning on CNNs and Vision Transformers

It is the removal of some model weights (setting them to zero) to reduce size and compute cost while preserving as much accuracy as possible.

Magnitude pruning only removes weights with small values, whereas the new method measures the geodesic distance in model space under the Fisher metric, reflecting the actual performance change rather than weight size alone.

Tests covered fully-connected networks and vision transformers on MNIST and CIFAR-10; large language models were not tested, so generalization remains unverified.

The results are early-stage research, but they open a path to cutting inference costs on constrained infrastructure, which is valuable for regional teams in the medium term.