Augmented data in Machine Learning : virtual sample generation

Details

Supervisors

Verleysen, Michel ; Lee, John Aldo

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique, à finalité spécialisée

Abstract

There is a growing interest from the private and public sectors in machine learning. These sectors will often face the problem of small data: they don't have enough data to properly feed and use most of the machine learning models. However, some workarounds exist such as transfer learning, which will use part of a model well trained to adapt it, or virtual sample generation, that will populate the small data set by generating new ones. These new samples are generated to correspond to the knowledge extracted from the small original data set, as if they were gathered by the same means. These generated samples, combined with the original ones, will increase the accuracy of the machine learning algorithms, allowing them to work in the same way as if the data set was large enough. Two different techniques, particle swarm optimisation virtual sample generation (PSOVSG) and differential evolution virtual sample generation (DEVSG), will be investigated, compared with each other as well as a baseline, to assess whether the model has improved thanks to these generated samples.