On Learning Wire-Length Efficient Neural Networks

Christopher Blake, Luyu Wang, Giuseppe Castiglione, Christopher Srinivasa, Marcus Brubaker,

Abstract: Pruning neural networks for wiring length efficiency is considered. Three techniques are proposed and experimentally tested: distance-based regularization, nested-rank pruning, and layer-by-layer bipartite matching. The first two algorithms are used in the training and pruning phases, respectively, and the third is used in the arranging neurons phase. Experiments show that distance-based regularization with weight based pruning tends to perform the best, with or without layer-by-layer bipartite matching. These results suggest that these techniques may be useful in creating neural networks for implementation in widely deployed specialized circuits.

TL;DR: Three new algorithms with ablation studies to prune neural network to optimize for wiring length, as opposed to number of remaining weights.

Keywords: energy, pruning, image recognition, neural networks