Abstract
With progress in the Internet of Things, artificial intelligence, and cloud/edge computing, urban sensing and computing (USC) has become a promising solution to address the challenges in modern cities. This article investigates how to combine the power of human/crowd and machine intelligence to enable more innovative USC applications.
| Original language | English |
|---|---|
| Pages (from-to) | 26-34 |
| Number of pages | 9 |
| Journal | Computer |
| Volume | 54 |
| Issue number | 4 |
| DOIs | |
| Publication status | Published - 9 Apr 2021 |
Bibliographical note
Funding Information:This work was supported by the National Natural Science Foundation of China under Grant 61872010.
Funding Information:
The evaluation of the existing research of USC tasks focuses mainly on metrics such as accuracy, preci sions, and recall. However, to enable city administrators’ and industry’s adoption of relevant techniques (for example, assisting in evidence-based policy making), the USC approaches should not be treated as black boxes. Instead, we should provide certain mechanisms, tools, or even platforms to make it easier for domain experts in urban monitoring, management, and planning to actively participate in the lifecycle of USC in an interactive way to achieve certain interpretability. Taking the PHM task in this article as an example, to make the inference model more interpretable, we need to incorporate state-of-the-art interpret ability techniques in machine learn ing with appropriate customization. For example, to achieve an instance- wise explanation on spatial correla tions extracted by deep learning models, we can use visualization to see if certain regions are closely related for some types of population health outcomes. Another commonly used approach is to compute a ranking of a set of features by techniques such as Local Interpretable Model-agnostic Explanations.15 Currently, the study yet been studied. For instance, users’ of human-in-the-loop interpretability sensitive information, such as locations has already drawn a lot of attention in and preferences, can be exposed during the domain of computer vision, but it the process of conducting CMH-USC has not yet been well studied in USC. tasks. To this end, how to protect users’ privacy and ensure that the CMH-USC Knowledge transfer system runs safely would be critical So far, CMH-USC tasks are conducted in practice. We believe that advanced mostly from scratch regarding the privacy protection and encryption data-collection process. That means schemes should be carefully considered that when we have a new task, we need in future CMH-USC application design. to collect the data from zero and then gradually build the applications. How- ever, there would be certain knowledge nspired by the complementary nat-in previous tasks that can help with the ure of human/crowd and MI, this new task, but a systematic approach is Iarticle investigated how to fully com-lacking to reuse such knowledge. For bine their power with coadaptation and example, suppose that we have already co-optimization in the context of USC. implemented an environment-sensing Specifically, we proposed a generic application in one city A, and now we framework for CMH-USC and demon-would like to build such an application strated two applications as case stud-for a new city B. Then, it is probable that ies. We also summarized a few research certain knowledge (for instance, some opportunities for the study of human– spatiotemporal correlations in the machine hybrid intelligence in USC environmental data) can be transferred applications. from city A to city B.18 In the future, determining how to ACKNOWLEDGMENT conduct knowledge transfer between This work was supported by the tasks would be a potential direction for National Natural Science Foundation CMH-USC research. Moreover, the idea of China under Grant 61872010. of transferring knowledge between cities is beyond the scope of CMH-USC, as this basic intuition may also benefit other smart-city applications such as planning and design. For instance, if we want to select locations for building certain facilities (say supermarkets) in a new target city B without historical data, we can discover the knowledge in city A, where we have a lot of data for planning, and transfer it to city B.
Publisher Copyright:
© 1970-2012 IEEE.
Copyright:
Copyright 2021 Elsevier B.V., All rights reserved.
ASJC Scopus subject areas
- General Computer Science