Intermittent Deployment of Branched CNN Models on Microcontrollers
Sanjita Bharat Salunkhe · 2023
Recent advances in the field of intermittent systems have led to a new branch of innovation where complex neural network models can be deployed on edge devices. With the increasing need to implement machine learning applications on small resource-constraint devices, it is important to find efficient networks and compatible intermittent systems. The main focus of this project will be to survey the various applications that have been introduced in recent years (discussed in section 3) and find an application that has not been deployed on a 16-bit microcontroller before. The next step will be deploying this application on the chosen intermittent system (BOBBER). Moreover, before the deployment, there are multiple stages/steps involved to make this application platform compatible which are described in section 4. Advanced techniques like quantization and quantization-aware training (QAT) will be used to reduce the size of the selected model, and their impact on the model’s size and accuracy will be analyzed. A detailed survey on the current applications will also help new researchers in the field to catch up on the work and advancements, and it could provide all the necessary details for choosing a particular model available in one place. The chosen network, SESR, will be discussed in detail and the section on the various available types of networks like branched and sequential will show the topology of the current advancements. The results and conclusion section 6 will discuss the outcome and the challenges that were faced during the experiments and conclude with ideas that could be further developed.