HEC Montreal

Published purpose

Deep learning models are increasingly being adopted and rapidly improved for property prediction in intelligent design, including predicting the functionalities of a novel circuit, foreseeing the working behaviors of a new material, and anticipating the reactions of a fresh compound. State-of-the-art deep techniques leverage two building blocks: deep graph neural networks and supervised signals from large amount of labelled data. Graphs are powerful and versatile data structures to model the relationships among objects, and powered by the graph structure, supervised learning leverages large amount of labelled data to construct efficient graph representation, which then enables the prediction model to make accurate decisions on the properties of a new design. Such state-of-the-art graph neural networks for property prediction, rely heavily on supervised learning with large labelled datasets, which is typically expensive and time-consuming. This Project will examine the development of fundamental techniques for graph representation learning in the presence of limited labeled data, aiming at accurate property prediction of novel designs. The successful outcome of this Project will be a learning framework that enables the efficient learning of graph representation with limited labelled data. Consequently, with improved graph representation, the properties of a new design can be more accurately and effectively evaluated, resulting in meaningfully speeding up the search for desirable designs.