Using a novel XAI algorithm for Data Augmentation in image classification problems

Abstract

Nowadays, Machine learning solutions are increasing their presence in the industry, and the benefits associated with the use of this technology are reflected in a reduction of time, costs, and a clear economic benefit, also most of these solutions use supervised learning, where a dataset with labeled real examples is needed.

 However, the many challenges associated with the implementation of a supervised machine learning solution are sometimes difficult to overpass, so one of the most important challenges is the creation of a big labeled dataset to feed the algorithms, since most of the time there is no any dataset available for the proposed solution, and most companies cannot afford to have a hand-created dataset with tens of thousands of records, also because they even cannot be sure that the model will work.

And the problem is that current machine learning techniques fail to improve the problem understanding on training with very small datasets, so here it is shown how by using a novel XAI method used to explain the decisions of a machine learning model in a computer vision problem, we can augment the labeled dataset focused on the important regions for the image classification, and increase the model performance, not just for training but for validation and testing, also this method turns out to be superior to the most used data augmentation methods if we further reduce the amount of information.