A team of researchers from the Data Science Institute at Columbia University (DSI) has received the Facebook Systems for Machine Learning Award in January 2020 for their research exploring how to co-train machine learning models not only for better decision-making but also to meet performance objectives.
The research has the potential to enhance machine learning models so that they are more scalable and adaptive.
"Today's machine-learning models are limited as they are primarily trained only for making the best predictions," says Asaf Cidon, an assistant professor of electrical engineering and computer science at Columbia and a Data Science Institute (DSI) member. He conducts this research with fellow DSI members and computer science professors Junfeng Yang and Suman Jana, whose combined expertise in computer systems and machine learning is ideally suited for the project.
"When deployed at scale, the models often face significant performance hurdles," says Cidon. "This project takes the first step in incorporating performance constraints and objectives in the process of model training."
As a first step, the team will focus on optimizing the performance and accuracy of recommendation systems. The systems currently used by Netflix for videos and Amazon for products, as well as Facebook's friend feed, are all examples of systems that could be enhanced by this research.
According to Cidon, today's recommendation systems encode each piece of content and users' past behavior as embeddings, or vectors that represent the content and how users interact with it. To compute a single recommendation, these systems must access in parallel hundreds of embeddings, and at a large scale, they must handle billions of daily recommendations.
The process is challenging from a performance perspective, since these embeddings are spread across hundreds or thousands of servers. The embeddings must also be read under strict deadlines, which are crucial in providing timely recommendations.
The team intends to design a recommender system that automatically learns the optimal layout and placement of the embeddings across thousands of servers. The system will be based on historical and dynamic access patterns to meet both performance objectives as well as to make the best possible recommendation for the user.