Building an intelligent, sustainable, and safe city is a global mission for both research communities and general public. Recent advance of technologies, such as Internet-of-Things (IoT), cloud computing, and machine learning, enables new solutions and applications, such as intelligent traffic system, urban environment monitoring, and smart energy systems. However, robust and intelligent network management on such a large scale becomes a great challenge. In this paper, we present a machine learning assisted network management framework for IoT in smart cities. This intelligent framework seamlessly integrates machine learning with sensing and communication, information fusion, and decision making in the city network architecture. We present a case study of vehicular sensing network for urban environment monitoring, in which the services are supported by various machine learning techniques. A novel distributed information fusion scheme that is robust to false data is then introduced. Finally we discuss scalable computation and implementation of machine learning methods in large scale networks.