Resource Trade-offs: Computation, Communication, and Information

Large-scale machine learning and data analysis is characterized by a significant consumption of heterogeneous resources, both computational and statistical. These include runtime, memory, communication in distributed applications, number of training examples, and more. Moreover, the required resources are often influenced by various constraints, such as maintaining data privacy, exhibiting robustness to outliers and noise, and tackling partial access to the data. Where traditionally these resources and constraints were studied separately, an emerging and recent paradigm is that they are all in fact correlated and can be studied in a unified framework. For example, it has been demonstrated that for some learning tasks, the availability of a larger training sample can have the following mutual effects: reduce the required runtime for learning the size of the memory utilized or the amount of communication, overcome partial information constraints, improve robustness, and generate outputs with privacy guarantees. As another example, analyzing the computational runtime required to achieve a certain statistical accuracy has proven to be a useful framework to study learning algorithms and their performance in a variety of settings.

The goal of this workshop is to bring together experts with diverse backgrounds whose work pertains to these questions and to exchange ideas while identifying new research opportunities for this emerging research area.