Energy Usage of Deep Learning in Smart Cities

Deep learning has increasingly become an essential component of many Smart City functions including smart city lightings, emergency rescues, smart drainage, and smart parking. These functions operate continuously in real-time throughout the day. Thus, excessive energy usage of deep learning computation can negatively impact economic benefits and efficiency of smart cities. The situation can escalate when dealing with resource-constrained large-scale smart cities of huge Internet-of-Things and networks with large numbers and varieties of sensors. To effectively sustain, manage and protect smart cities from failures due to energy overload, the awareness of energy consumption by deep learning computation is unavoidably necessary. Most recent research in smart cities focuses on using deep learning to perform certain tasks but does not address energy issues. This paper presents a formal approach to estimating energy consumption of deep learning and illustrates its use in smart cities. In particular, we develop a fine-grained mathematical model that extends an existing model to include the quantification of MAC (multiply-and-accumulate) operations as well as data access from a memory hierarchy. This paper focuses on deep and convolutional neural networks. We describe the proposed approach and validate the results obtained from our model by comparing them against those of existing work. The proposed approach is applied to three (deep) neural systems in smart cities, namely smart drainage, smart transportation and smart parking systems, all of which yield promising results.