Machine learning is a data-driven science. It generates computer programs from data to perform tasks, without being explicitly programmed to do so. The Graph Machine Learning group researches machine learning approaches for relational and geometrical data using GraphAI technology developed by us at NEC Laboratories Europe. We focus on exploring new ways we can apply machine learning that challenges standard (deep) learning methods and advances society.
Advancing relational learning with GraphAI technology
Our team researches graph-structured, relational learning. Using GraphAI, we analyze relational or graph-structured data that represents the features of data points, their relationships and interdependencies.
GraphAI helps overcome the problem of missing data points in complex data dependencies, meeting the needs of emerging use cases.
We address various issues of using machine learning for size of errors in its own predictions. Our approach integrates probabilistic modeling with deep graph networking to robustly model the propagation of graph uncertainties.
In addition to relational learning, we research geometric deep learning to advance technologies in computational science. With this research, we model the interaction of physical entities in a certain range of space. Conventional numerical simulation requires huge computing resources. By replacing simulation with machine learning, and improving on existing approaches, we are reducing the number and duration of scientific tasks needed for research. For example, we recently developed a surrogate model to solve partial differential equations that drastically reduces the computation load and cost of machine learning. With a sufficient amount of data, the model can even predict future changes to a physical system with no knowledge of its partial differential equations.
Discovering new materials with materials informatics
Recent applications of GraphAI technology have focused on the materials science industry. A NEC study of related use cases shows how leveraging GraphAI’s relational learning and geometric deep learning methodologies will help achieve a digital transformation in materials science and the chemical industry. With relational learning, knowledge between scientific documents and a product developer’s requirements can be linked. Geometric learning augments simulations such as molecular dynamics and ab initio theory–based calculations.
Through our research activities, the Graph Machine Learning group explores the frontiers of machine learning applications, helping find solutions to some of society’s most pressing technical challenges. We conduct research in computer science and numerous industry domains, including public services, manufacturing, finance, biomedical, physics, and chemistry.