Deep Networking for IoT Modules
Аннотация
Deep neural network (DNN) based internet of things (IoT) devices are expected to change a broad spectrum of industrial applications, backed by ecosystems that have been successfully identified over the last few years. DNNs, on the contrary, learn from a large amount of data that is generated by IoT devices and consist of various parameters and processes which these DNN will take to. This causes high power consumption and delays in data processing. Various new strategies are on the horizon to combat these issues and make DNNs capable of functioning in real-time within resource constrained IoT devices. In this research, we investigate the existing hardware-software co-design techniques proposed to be used on resource-constrained edge devices for running DNNs. In this paper, we explore these trade-offs in model size vs. classification accuracy and deploy time/energy consumption. Overview of DNNs Let us take an example to explain this. This is followed by an overview of a number tools to enable running DNNs on resource-constrained hardware platforms. We then illustrate the concepts of memory hierarchy, dataflow mapping techniques. Furthermore, some model optimization techniques like quantization and pruning are discussed. We also present several case examples to illustrate the practicality of deploying DNNs in IoT applications. Finally, we briefly explain the results and unmet research needs specific to each design aspect in addition to recommandations for further work. The proposed review could guide the design and roll-out of subsequent software- and hardware-native solutions to practical IoT applications.