Professors Asaf Cidon, Suman Jana and Junfeng Yang have been awarded with Facebook's Systems for Machine Learning research award, for their proposal on building machine learning systems that co-trained not only for making better decisions, but also for meeting performance objectives.
The proposal focuses on deep learning recommender systems, which power many popular online services, such as Facebook's friend feed, and Netflix's video and Amazon's product recommendations. These systems encode each piece of content (e.g., video or post) and user as embeddings, which are vectors that represent the content and how users interact with it. To compute a single recommendation, these systems need to access hundreds of embeddings in parallel. This process is very challenging from a performance perspective, since these lookups are spread across hundreds or thousands of servers, and they have to be completed under stringent deadlines since they sit in the critical path of providing the recommendation. Today the layout and placement of these embeddings is static and determined manually.
The Columbia researchers plan to design a recommender system that automatically learns the optimal layout and placement of the embeddings across thousands of servers, based on historical and ongoing access patterns, in order to meet both performance objectives as well as to make the best possible recommendation for the user.
According to Professor Cidon, "Historically, machine learning models have been trained solely for making the best prediction. However, often times models and the data they rely on do not scale when they are deployed in large-scale production environments. This proposal takes a first and important step of co-training both for prediction accuracy, as well as for meeting performance objectives constraints. Facebook's support is crucial for the success of this line of research, because their models run at a truly global scale, which is hard for us to emulate on our own in a university setting."