CAREER: Advancing the Frontier in System Architectures for Artificially Intelligent Services and Applications

点击翻译按钮获取中文摘要

The way society interacts with today's mobile technologies is rapidly changing as such devices - omnipresent in our lives - become increasingly personal and knowledgeable. Intelligent assistants (IAs), such as Apple's Siri, Google's Google Now, Microsoft's Cortana, and Amazon's Echo, are continuing to become more sophisticated and are expected to grow in popularity as wearables continue to gain traction. The computing capabilities required to support these technology trends continue to evolve toward intelligent systems that use sophisticated machine learning and computer vision algorithms on the critical path of audio and vision interactions. In this proposed work, PI Jason Mars aims to understand how future cloud and mobile systems should be designed (and co-designed) to support increasing demand from users. This proposed work advances both 1) the efficiency of current computing platforms to reduce their energy, cost, and environmental footprint, and 2) the expansion of open end-to-end systems with state-of-the-art algorithmic capabilities to enable new, more sophisticated, intelligent technologies and usher in future research and inquiry. PI Mars' technical approach incorporates three pillars of innovation. These pillars span application insights, to inform the design of underlying systems, vertical innovation to advance the cross-layer design of the system stack for emerging applications and services, and horizontal innovation to reason about a computational fabric spanning mobile hardware and cloud hardware to create a unified platform for computation. In addition to having impact on national interests, economic advancement, and technology in general, this proposed work incorporates significant innovation in undergraduate, and graduate education. The Intelleco and DeepSirius systems described in this proposal will be used to design an undergraduate and graduate course for state-of-the-art datacenter design. In addition all artifacts and teaching materials will be disseminated broadly through open source and creative commons.