A Dual-Stream Recurrent Neural Network for Student Feedback Prediction using Kinect

Convenience internet access and ubiquitous computing have opened up new avenues for learning and teaching. They are now no longer confined to the classroom walls, but are available to anyone connected to the internet. E-learning has opened massive opportunities for learners who otherwise would have been constrained due to geographical distances, time and/or cost factors. It has revolutionized the learning methods and represents a paradigm shift from traditional learning methods. However, despite all its advantages, e-learning is not without its own shortcomings. Understanding the effectiveness of a teaching strategy through learner feedback has been a key performance measure and decision making criteria to fine tune the teaching strategy. However, traditional methods of collecting learner feedback are inadequate in a geographically distributed, virtual setup of the e-learning environment. Innovative and novel learner feedback collection mechanism is hence the need of the hour. In this work, we design and develop a deep learning based student feedback prediction system by recognizing the subtle facial motions during a student's learning activity. This allows the system to infer the needs of the learners as if it is a real-human teacher in order to provide the appropriate feedback. We propose a recurrent convolutional neural network structure to understand the color and depth streams of video taken by an RGB-D camera. Experimental results have shown that our system achieve high accuracy in estimating the feedback labels. While we demonstrate the proposed framework in an e-learning setup, it can be adapted to other applications such as in-house patient monitoring and rehabilitation training.