Deep Augmentation in Convolutional Neural Network

Abstract

Deep Convolutional Neural Networks (ConvNets) have been tremendously successful in the field of computer vision, particularly in image classification and object recognition tasks. A huge amount of training data, computing power and training time are needed to train such networks. Data augmentation is a technique commonly used to help address the data scarcity problem. Although augmentation helps with the problem of training data scarcity to some extent, huge amounts of computing power and training time are still needed. In this work, we propose a novel approach to training deep ConvNets, which reduces the need for huge computing power and training time. In our approach, we move data augmentation deep in the network and perform augmentation of high-level features rather than raw input data. Our experiment shows that performing augmentation in feature space reduces the training time and even the computing power. Moreover, we can break down our model into two parts (pre-argumentation and post-augmentation) and train one part at a time. This allows us to train a bigger model in a system than it could normally handle.