Detecting educational emotion of students is important as this plays a vital role in their learning process. Generally, in a regular classroom, instructors can observe the emotion of the students by their facial expressions. In an online learning platform, it is quite challenging. Deep learning architectures are found to be efficient in detecting emotion from facial expressions. However, these architectures are very deep in nature and computationally expensive, which are not suitable to deploy on students' edge devices. In this study, we propose a deep learning architecture based on MobileNet, which is lightweight in nature and suitable to deploy in edge devices. We performed a comparative analysis of the proposed architecture with some other state-of-the-art deep learning architectures using a dataset called "Spontaneous Facial Expression Database for Academic Emotion Inference in Online Learning (OL-SFED)"which was developed using an online learning platform. From the comparison, we found that the proposed architecture showed competitive performance in terms of accuracy with the state-of-the-art architectures while using a significantly less number of parameters than the others.