The architectures underlying modern network hardware and software have their roots in designs that were developed decades ago. Even though these architectures have evolved in many ways over the years, they remain unchanged in two key aspects: (1) They support ?exact? or complete/absolute reliable communication (either hop-by-hop, or end-to-end, or both); (2) They adhere to strict layering, and the resulting encapsulation and interfaces hide from lower network layers the semantics of the data applications transmit over the network.
These design principles place serious impediments for emerging distributed machine learning (ML) training and inference applications. These applications are seeing adoption in a wide variety of important domains, such as, computer vision, robotics, data science, graphics, and speech recognition. Two distinguishing attributes of these applications are: (1) their computations are intrinsically inexact in nature, because these applications rely on computing or utilizing statistical models, and (2) their input and intermediate data have well-defined structure, i.e., tensors, or multi-dimensional arrays of typed data. Give these attributes, enforcing exact communication in a data semantics-unaware fashion limits the potentially enormous benefits of embracing inexactness in these approximate applications.
This project explores co-designing ML applications with layers of the network software and hardware stack to allow application-driven cross-layer optimization for energy efficiency, hardware density/capacity, and performance. Given an application-provided overall inexactness budget, this research will explore both how to systematically apportion the budget across network layers, and how different layers can reconfigure their functionality to achieve different levels of approximation.
This project will develop strawman approaches to encoding structured data and to achieving budget-driven inexact computation over it. The research will use experiments, simulations, and analysis to identify performance benefits to ML applications, and fundamental trade-offs that determine the feasibility of this approach. The resulting inexactness-aware ML software stack could drive hitherto unseen performance and accuracy improvements, and potentially drive future innovations in ML algorithms, systems, and applications.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
