Parameterized DNN design for identifying the resource limitations of edge deep learning hardware

Abstract

Artificial intelligence has come a long way in the last several years and has made remarkable advancements, with deep learning neural networks emerging as a powerful tool for solving complex problems. Even with the technology having rapidly grown to such incredible levels thus far, one of the most significant challenges still standing is how to effectively deploy deep learning algorithms on edge devices such as mobile phones and Internet of Things (IOT) devices while preserving performance. Edge devices often face limitations in resources, such as memory and processing power, which can raise challenges for running complex deep-learning models. The fundamental goal of this project is to evaluate the resource constraints of edge devices, allowing for the deployment of deep learning algorithms without sacrificing performance. This enables tasks such as image recognition and speech processing to be carried out directly on edge devices, minimizing the need to rely on powerful cloud servers. In this final year project (FYP), a program to identify resource limitations on edge learning devices while inferencing deep learning models will be developed using Python and the deep learning framework PyTorch. Furthermore, the program will tune hyperparameters such as input shapes and batch sizes, as well as model complexity such as feature maps and weights, to achieve an ideal state that minimizes resource demands while preserving performance.