Synthetic Data Generation for Machine Learning

Training machine learning models often requires a very large amount of data. Although we generate millions of terabytes of data every day, there isn’t enough data to accurately address many tasks. Collecting and organizing large volumes of data for each type of application is already a considerable challenge for many individuals and organizations. However, whether for technical or ethical reasons, to conduct research and create artificial intelligence applications, it is necessary to take extra care in the construction of these datasets, such as categorizing and balancing the samples and respecting people’s privacy.

Using synthetic data for model training can be a solution to overcome several of the problems related to constructing datasets. Recent studies point to the possibility of using synthetic data to train models that perform well when applied to real data. In this series of three seminars, we will present: advantages and limitations of this method, in particular, in the areas of computer vision and computer graphics; how synthetic data can be generated on a large scale using the computer graphics tools available today; and how this idea expands to machine learning within simulations.