PI: Chi-Sheng Shih (National Taiwan University), CoPI: Kwei-Jay Lin (University of California, Irvine)

Champion: Srenivas Varadarajan (Intel)

Status Quo: 

According to the report of IDC (International Data Corporation) [1][2], by 2019, at least 40% of IoT-created data will be stored, processed, analyzed, and acted upon close to, or at the edge of, the network. The edge devices such as machines (sensors), mobile devices, and computers will convert disparate data into meaningful information. With derived information, services can be provided in many scenarios and real-time events. 

However, given the limited processing capability on most edge/sensor devices, as well as potentially dynamic/unpredictable application environments, we need a flexible end-to-end distributed IoT computing architecture to handle computation-intensive tasks such as object recognition and localization in real time.

For example, in the augmented collected beings scenario, although mobile devices can act as augmented sensing devices for human, these devices usually have limited computing capabilities, and cannot fully process sensing data to provide promptly response to the users. For another example, in the robotic scenarios, image sensors on drones/robots can take images of the landmarks, objects and people in a target environment, but they may not have sufficient computing capabilities to correctly recognize landmarks and people using solely local computation resources under different context and lighting conditions. 

Key New Insights:

We argue that one of the main challenges for augmented sensing is to provide timely intelligence under timing varying context where IoT systems must dynamically and intelligently distribute the computation workloads among sensing, edge and cloud devices to trade off real-time response and system intelligence. To achieve this, we propose a flexible DEC (device-edge-cloud) middleware a.k.a., DynaCollect DEC. This middleware will build an intelligent performance-driven edge-based framework to dynamically distribute semi-supervised (deep) learning workloads in order to support smart IoT applications.

Machine learning approaches have shown promising results for object recognition in limited use scenario. However, it remains open for using single model to recognize objects under different brightness, background, and geographical locations. Moreover, it is still challenging for many sensing devices and edge devices to support such computation workload in real-time. To tackle this issue, the sensing devices, edge devices, and cloud devices have to collaborate in a dynamic manner to share the workloads. 

Our preliminary studies show that the above goals can be achieved by (1) re-training the model for different context from time to time and (2) dynamically distributing computation workloads among devices for different context. Our studies also show that the above two operations cannot be conducted on sensing devices in general. In this project, we will develop the workload distribution mechanisms by learning the performance requirement, device properties, and system context from time to time. 



[1] IDC's Worldwide IoT Infrastructure Taxonomy, 2017

[2] IDC FutureScape: Worldwide Internet of Things 2017 Predictions

 (Updated in Aug, 2017)