CSR: Small: ARTEMIS: Algorithm-Hardware Co-Design for Efficient Machine Learning Systems

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

With the increased popularity of machine learning algorithms deployed on a variety of hardware systems, the problem of identifying the best model among numerous possible configurations has drawn significant attention. The problem is compounded by the need to select the right platform to run these applications, under given power or latency constraints. This "hardware wall" forces machine learning service providers to constantly redesign the underlying hardware fabric to satisfy certain constraints. This project develops tools for automatic and efficient co-design of machine learning algorithms and hardware platforms that will result in significant cost and time-to-market reduction for machine learning systems. The project introduces efficient meta-learning for machine learning systems and algorithm-hardware platform co-design. Specifically, the project will develop meta-learning algorithms for the optimization of machine learning models under system hardware constraints and formulate the hardware design of efficient machine learning systems as a machine learning problem itself, that can be effectively solved by meta-learning optimization algorithms. Finally, the project will develop multi-objective algorithms for the co-design of machine learning applications and hardware platforms they need to run on, and exploit domain knowledge from hardware engineering and design schemes to substantially accelerate hardware-aware model optimization. The results of the project seek to change the landscape of modeling, optimization, and design methodologies for efficient machine learning systems. Furthermore, the work aims to have an important educational and mentoring component by potentially changing how engineers are trained in a multidisciplinary fashion for dealing with next generation technological advances in general, and the problem of efficiently and intelligently co-designing machine learning algorithms and the hardware platforms they are running on, in particular. The project will involve a diverse graduate and undergraduate trainee population, while expanding the project's outreach to high-school and middle-school students. The data, code, results, and simulators developed in this project will be made available publicly throughout the duration of the project and for at least four years after the end of the project. The location of the repository is on the website of Carnegie Mellon University's Energy Aware Computing group (www.ece.cmu.edu/~enyac). 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.