Abstract
Over the past decade, the use of machine learning has increasedexponentially. Models are far more complex than ever before, growing togargantuan sizes and housing millions of weights. Unfortunately, the fact thatlarge models have become the state of the art means that it often costsmillions of dollars to train and operate them. These expenses not only hurtcompanies but also bar non-wealthy individuals from contributing to newdevelopments and force consumers to pay greater prices for AI. Current methodsused to prune models, such as iterative magnitude pruning, have shown greataccuracy but require an iterative training sequence that is incrediblycomputationally and environmentally taxing. To solve this problem, BINGO isintroduced. BINGO, during the training pass, studies specific subsets of aneural network one at a time to gauge how significant of a role each weightplays in contributing to a network's accuracy. By the time training is done,BINGO generates a significance score for each weight, allowing forinsignificant weights to be pruned in one shot. BINGO provides anaccuracy-preserving pruning technique that is less computationally intensivethan current methods, allowing for a world where AI growth does not have tomean model growth, as well.