A Novel Approach Using MPI to Reduce Training Time for Capsule Networks
David Schwab, Yang Li · 2019
The deep learning environment is currently developing at an exciting pace as new techniques and new results are consistently being released through new research. One such new development is the capsule network which is a novel new approach to deep learning wherein the standard neurons of a neural network are replaced with capsules. Capsules differ from neurons in that they receive vectors as input and create vectors at their output. Capsule networks additionally are good at classifying objects irrelative of the orientation of the object in a given image. Capsule networks can also train on less data than traditional convolutional neural networks (CNNs). In this paper we present an experiment wherein we aim to reduce the training time of capsule networks, while still retaining acceptable accuracy. We accomplish this by dividing the training set among multiple nodes in a cluster environment and using MPI to allow nodes to communicate their classification results to vote on the correct classification for each image.