Software Library for Generative Model Applications

Abstract

The generation of data by machine learning models is a powerful concept that has impacted the field of Artificial Intelligence in the past few years. In this thesis, we focus on building a software library to facilitate the workflow, evaluation, and analysis of generative models. Our work is primarily aimed at helping a specialty chemicals company use a state of the art molecule generation model for their specific applications. We reference the body of work containing the model as DEG, short for Data-Efficient Graph Grammar Learning for Molecular Generation [16]. DEG is capable of creating synthesizable molecules from small amounts of data, making it quite attractive for companies looking for practical methods to explore new molecules. As an overarching goal, we will design our library to incorporate other types of generative models and become a tool that the field can benefit from.