CAREER: An Automated End-to-end Machine Learning System

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

Machine learning (ML) has surpassed human performance in some contexts, including image classification, natural language processing, game playing, and content generation. ML’s success is enabled by the recent development of ML systems that offer high-level programming interfaces for people to prototype various ML models on modern hardware platforms. However, deploying these models in diverse, real-world computing environments requires significant engineering effort to design and implement the required performance optimizations. To address this challenge, this project explores an automated, end-to-end approach to building efficient, scalable, and sustainable ML systems for diverse ML applications and hardware platforms. The project takes a bottom-up, three-pronged approach that involves (1) automatically discovering and verifying various systems optimizations for different ML models and hardware backends; (2) new methodologies for applying the discovered systems optimizations in an end-to-end fashion; and (3) combining systems and ML optimizations for fast and accurate ML computations. As ML techniques move closer to end-users and become increasingly integrated into today’s society, the proposed work can effectively reduce the energy consumption and financial cost of modern ML techniques. The key improvements of the proposed research will arise along two axes: (1) replacing manually designed performance optimizations used in today’s ML systems with automated generation, verification, and application of systems optimizations for ML computations on modern hardware platforms; and (2) democratizing ML techniques by lowering the monetary cost of developing and deploying ML applications. The project also includes outreach activities to attract students from populations currently underrepresented in computing. Key to these activities is embracing the interdisciplinary nature of ML systems research, which spans computer systems, compilers, programming languages, and machine learning. The software artifacts of this project will be released and regularly maintained. 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.