Urban Intelligence with Deep Edges

Abstract

With the increased accuracy available from state of the art deep learning models and new embedded devices at the edge of the network capable of running and updating these models there is potential for urban intelligence at the edge of the network. The physical proximity of these edge devices will allow for intelligent reasoning one hop away from data generation. This will allow a range of modern urban reasoning applications that require reduced latency and jitter such as remote surgery, vehicle collision detection and augmented reality. The traffic flow from IoT devices to the cloud will also be reduced as with the increased accuracy from deep learning models only a subset of the data will need to be reported after a first pass analysis. However, the training time of deep learning models can be long, taking weeks on multiple desktop GPUs for large datasets. In this paper we show how transfer learning can be used to update the last layers of pre-trained models at the edge of the network, dramatically reducing the training time and allowing the model to perform new tasks without data ever having to be sent to the cloud. This will also improve the users' privacy, which is a key requirement for urban intelligence applications with the introduction of GDPR. We compare our approach to alternative IoT urban intelligence architectures such as cloud-based architectures and deep learning algorithms trained only on local data.