Content area
Abstract
Cloud computing is the primary carrier of artificial intelligence services and deep learning algorithms by providing powerful computation and storage resources. However, the device-cloud workflow makes it difficult to satisfy both computation-intensive and time-intensive requirements. Moreover, it is challenging for cloud computing to perform privacy-sensitive applications. To address these problems, we focus on providing edge systems support for intelligent services and algorithms. By migrating intelligence from the cloud to the edge to enable cloud-edge collaboration, we improved the quality of experience of intelligence services with faster response speed and shorted communication paths. By carefully designing the deep learning algorithm for edge-edge collaboration, we improved the data-scarce issue widely existing in the healthcare field with the protection of data privacy.
Our solution starts by profiling the cloud-based service performance to analyze the drawbacks of the cloud-only approach for intelligent services and using EdgeBroker to improve the user experience with online learning. EdgeBroker can improve the quality of experience by 1.5x in a real-world deployment. We then migrated the intelligence from the cloud to the edge to break the integrated device-cloud workflow. We proposed HomeCache, which is a voice caching system designed for home automation. It can improve the inference speed by 70% and keep low resource consumption on resource-constraint edge hardware. We also discovered that natural language-based intelligent services have precise redundancy and fuzzy redundancy. We proposed a general and scalable edge caching system SemCache, improving the inference speed for natural language-based intelligence services in multiple target scenarios by around 91.7% and 81.6% for audio and text requests, respectively. To address the data privacy problem, we selected the healthcare domain to develop DREAMS. This distributed deep reinforcement learning system can address the scarce annotation data problem, protect data privacy, and enable multi-institutional collaboration with the trade-off of time efficiency.
