Preservation of Higher Accuracy Computing in Resource-Constrained Devices Using Deep Neural Approach
R. Manikandan, T. Mathumathi, C. Ramesh, S. Arun, R. Krishnamoorthy, S. Padmapriya · 2022 Second International Conference on Artificial Intelligence and Smart Energy (ICAIS) · 2022
The embedded type of de vices in IOT generally depends upon the resource constraints which include memory capabilities, low power consumption and reliable in cost. The constrained devices such as edge server are handled at the end nodes. The end nodes such as sensors and actuators are connected using the gateway devices which connect the IOT cloud-based platform. A wireless device which has the limited set of processing and storage-based capability which runs based on the wireless medium or batteries is the resource constrained device. Resource constrained devices provides the efficient way of limited processing with the maximal data output along with the minimal power as input. These are generally cost effective as it consumes less energy and power consumption in devices. The edge server is a type of resource-constrained devices which is the entry point of the network and application. In this paper, the research is based upon the proposal model of resource constrained devices by reducing the parameters using the DNN. The DNN model parameters reduce the memory, execution latency by attaining the higher accuracy. To preserve the higher accuracy in the device computation, the Knowledge Distillation Method is proposed. The knowledge distillation method determines the output predictions of larger DNN into the smaller DNN trained sets. This methodology reduces the trained model by compressing the model accordingly. These smaller DNN predicts the output and behaviors similar to the larger DNN. Smaller DNN predicts approximately equal to the larger DNN. Knowledge Distillation Method is used in several applications in machine learning such as natural language processing, AI, Detection of objects and neural networks graph correspondingly.