An Empirical Study on Improving the Speed and Generalization of Neural Networks Using a Parallel Circuit Approach

摘要

One of the common problems of neural networks, especially those with many layers, consists of their lengthy training time. We attempt to solve this problem at the algorithmic level, proposing a simple parallel design which is inspired by the parallel circuits found in the human retina. To avoid large matrix calculations, we split the original network vertically into parallel circuits and let the backpropagation algorithm flow in each subnetwork independently. Experimental results have shown the speed advantage of the proposed approach but also point out that this advantage is affected by multiple dependencies. The results also suggest that parallel circuits improve the generalization ability of neural networks presumably due to automatic problem decomposition. By studying network sparsity, we partly justified this theory and proposed a potential method for improving the design.