AI in the Wild : Robust evaluation and optimized fine-tuning of machine learning algorithms deployed on the edge

In research on applying Machine learning to Embedded systems in the field, little research has been done in the evaluation of deployed devices. Often a system is trained with lots of data, hoping to encapsulate everything that a deployed system may encounter. Next to that, real-life data encapsulates the need to research data which has not been seen yet and thus research the influence of out-of-distribution data. In this research, attention will be brought to the evaluation of deployed edge devices and the trade-off between evaluation reliability and power consumption. Furthermore, the trade-off between using cloud hardware and edge hardware is discussed, where it finally culminates in the trade-off between fine-tuning on the edge versus fine-tuning in the cloud. These questions should give insight into the evaluation of a deployed system using edge intelligence, as well as how best to train and further fine-tune these systems