With the emergence of the edge computing paradigm, many edge applications, such as image recognition and augmented reality requires performing machine learning and artificial intelligence workloads on edge devices. Most ML models are large and computationally heavy, whereas edge devices are usually equipped with limited power and energy. Unfortunately, small models cannot perform well. Recent works use KT technique to improve performance of smaller models. The research teams results show that KT technique behaviors and outcomes differ from one architecture to another, and the effectiveness of KT technique depends on the training dataset and network architecture.