CityFlow: Supporting Spatial-Temporal Edge Computing for Urban Machine Learning Applications

Abstract

A growing trend in smart cities is the use of machine learning techniques to gather city data, formulate learning tasks and models, and use these to develop solutions to city problems. However, although these processes are sufficient for theoretical experiments, they often fail when they meet the reality of city data and processes, which by their very nature are highly distributed, heterogeneous, and exhibit high degrees of spatial and temporal variance. In order to address those problems, we have designed and implemented an integrated development environment called CityFlow that supports developing machine learning applications. With CityFlow, we can develop, deploy, and maintain machine learning applications easily by using an intuitive data flow model. To verify our approach, we conducted two case studies: deploying a road damage detection application to help monitor transport infrastructure and an automatic labeling application in support of a participatory sensing application. These applications show both the generic applicability of our approach, and its ease of use; both critical if we wish to deploy sophisticated ML based applications to smart cities.

Publication
Urb-IoT 2018: 3rd EAI International Conference on IoT in Urban Space pp 3-15
Makoto Kawano
Researcher

My research interests include efficient machine learning for real-world deployment.