One-Shot Pruning Beats Iterative Pruning in a Compact Traffic-Flow MLP
Arunya Arunya
Abstract
A plain MLP, pruned to 80\% sparsity and quantized to int8, matches the best published $R^2$ on the Huawei intersection dataset and stays competitive on Metro Interstate at a third of the storage. Iterative magnitude pruning, the standard refinement over one-shot, is consistently worse at this scale.
Chat is not available.
Successful Page Load