Infor Partnering with Oxford and Carnegie Mellon on Machine Learning
Infor awarded an unrestricted research gift of $100,000 USD to support on-going work on in-database machine learning to the Factorised Databases (FDB) Project of the Computer Science Department at Oxford in the UK.
Professor Dan Olteanu leads the FDB Project. “Our goal is to build a scalable system for training machine learning models over relational databases. Our approach comes with both theoretical and practical benefits. It enjoys lower computational complexity than the existing approaches, which means, in practice, training over larger datasets and orders-of-magnitude faster than state-of-the-art analytics systems.”
Olteanu previously consulted on development of Infor’s LogicBlox. This new work is expected to advance Infor’s efforts to move from machine learning to deep learning on LogicBlox, a core part of Infor Retail’s demand forecasting, allocation, replenishment, and assortment optimization capabilities.
“Infor is pleased to support the work of Professor Olteanu on the FDB Project, which addresses a fundamental obstacle to scaling machine learning to large data sets, such as we encounter when forecasting demand in retail contexts,” said Kurt Stirewalt, Infor VP of software development.
“This research team has developed and continues to research methods to do factorized machine learning ‘in the database’ to avoid the biggest bottleneck in the process, he added. “The results from the Oxford team show this can be done in tens of seconds rather than hours or days, which represents the current standard.”
Infor began a similar partnership with Carnegie Mellon University in Pittsburgh, Pennsylvania, in October. Infor gave an unrestricted gift of $100,000 to Professor Benjamin Moseley for his work on advanced algorithms for scheduling execution graphs of declarative languages. This research is expected to improve the task scheduler of the LogicBlox platform.
“LogicBlox currently uses a very expensive algorithm to do this scheduling, and we have long believed that a more efficient algorithm could provide substantial speed-up to both existing and new applications with little or no application-level tuning required,” Stirewalt says.