Fine Granularity is Critical for Intelligent Neural Network Pruning

Neural network pruning is a popular approach to reducing the computational costs of training and/or deploying a network, and aims to do so while minimizing accuracy loss. Pruning methods that remove individual weights (fine granularity) yield better ratios of accuracy to parameter count, while methods that preserve some or all of a network’s structure (coarser granularity, e.g. pruning channels from a CNN) take better advantage of hardware and software optimized for dense matrix computations. We compare intelligent iterative pruning using several different criteria sampled from the literature against random pruning at initialization across multiple granularities on two different image classification architectures and tasks. We find that the advantage of intelligent pruning (with any criterion) over random pruning decreases dramatically as granularity becomes coarser. Our results suggest that, compared to coarse pruning, fine pruning combined with efficient implementation of the resulting networks is a more promising direction for improving accuracy-to-cost ratios.