Understanding data augmentation for classification: when to warp?

TL;DR

This paper compares data augmentation techniques in data-space and feature-space for image classification, finding data-space augmentation more effective when plausible transformations are known.

Contribution

It provides an empirical evaluation of data warping versus synthetic over-sampling for augmenting training data in neural networks and SVMs.

Findings

Data-space augmentation yields better performance when plausible transformations are available.

Augmentation reduces overfitting across different classifiers.

Feature-space augmentation is less effective without suitable transformations.

Abstract

In this paper we investigate the benefit of augmenting data with synthetically created samples when training a machine learning classifier. Two approaches for creating additional training samples are data warping, which generates additional samples through transformations applied in the data-space, and synthetic over-sampling, which creates additional samples in feature-space. We experimentally evaluate the benefits of data augmentation for a convolutional backpropagation-trained neural network, a convolutional support vector machine and a convolutional extreme learning machine classifier, using the standard MNIST handwritten digit dataset. We found that while it is possible to perform generic augmentation in feature-space, if plausible transforms for the data are known then augmentation in data-space provides a greater benefit for improving performance and reducing overfitting.