Light weight AI models for low-power edge devices

Abstract

Edge computing has evolved as a promising area of AI, where the processing of data occurs near the source of the data instead of relying on cloud computing. The advantages of edge computing are that it helps reduce latency, increase the speed of responses, and maintain the privacy of the data. However, deep learning models require a high amount of computational resources, memory, and power, making them less suitable for edge devices such as IoT devices, sensors, and mobile devices. To overcome these limitations, a number of lightweight AI models have been developed that require less computational resources and provide acceptable accuracy. This paper focuses on light-weight neural network architectures and optimization techniques. The paper also discusses a number of model compression techniques, such as pruning and quantization, that help optimize the performance of the model and reduce the resources required for the execution of the model. Tiny ML helps deploy machine learning models on microcontroller devices and requires minimal power for execution. This paper also focuses on light-weight neural network architectures such as Mobile Net and Efficient Net, where efficient performance is achieved through efficient design strategies. In addition, this paper focuses on knowledge distillation as a technique that helps map complex models onto light-weight models. Edge intelligence helps perform real-time operations and maintain the privacy of the user. A light-weight model approach for efficient performance and minimal power consumption of edge devices has been proposed. The results of the study show that light-weight AI models can be deployed on edge devices with minimal loss of accuracy, making them efficient for realtime operations.