CNS Core: Medium: Accurate Anytime Learning for Energy andTimeliness in Software Systems

University Of Chicago, Chicago IL

Investigators

Abstract

Modern software systems increasingly rely on deep neural networks to perform a wide range of tasks, such as natural language translation and autonomous driving. The key to their success is that deep neural networks can well approximate these difficult tasks. Unfortunately, the more accurate the approximation, the more resources required. When deployed on mobile devices or autonomous vehicles those resource needs directly impact people as the time or energy/battery required to produce an answer. This project tackles this crucial problem by developing sound engineering methods to make disciplined tradeoffs between neural network accuracy and resource usage in software systems. The proposed work will take an interdisciplinary approach. Its first thrust will design novel neural networks that efficiently produce a series of outputs, instead of just a single output, such that the output accuracy will increase with increasing resources. The second thrust will develop new resource management software that automatically adjusts both underlying system settings and one or multiple neural networks to meet high-level software accuracy and energy requirements. The third will create tools for automatically analyzing the software context where neural networks are used, inferring accuracy and resource requirements for neural networks and identifying inefficient use of neural networks. The project has the potential to improve the efficiency and reliability of software that incorporates neural networks, and hence improve people's daily life experiences. Software developers will benefit from greater flexibility in neural network design, greater assurance that the neural networks they deploy will meet their accuracy and resource requirements, and greater understanding of how the neural network impacts the rest of their software systems. Additionally, the project will create many educational opportunities through enhanced classroom projects and creation of research opportunities for undergraduates, broadening the participation in computing. All of the data, code, results, and artifacts of this project will be made publicly available through the webpage https://alert.cs.uchicago.edu/. They will be available on-line for a period of at least five years following the completion of this project. 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.