Lightweight and Interpretable Detection of Affective Engagement for Online Learners

Offering online learners with a personalized, ethical, and privacy-protecting pedagogical companion that can detect their affective engagement and that can run on the devices they use to learn (edge devices) requires not only generating lightweight models but also interpretable, trustworthy, and generalizable ones. SqueezeNet is a lightweight deep neural network architecture, which is more popular for the edge devices. This paper showcases how SqueezeNet appears not to provide as meaningful explanations of its predictions of academic emotions as those of a lightweight traditional CNN model. More specifically, the winning CNN model, consisting of 473,317 parameters, delivers a predictive accuracy of 99.4% on the testing set, while maintaining an identical level of descriptive accuracy. In contrast, the best SqueezeNet model comprises 141,301 parameters (3x less) and has a predictive accuracy of 93.7% and a dropping descriptive accuracy of 84.0%. In brief, the SqueezeNet model does not maintain a similarly high descriptive accuracy and does not effectively identify facial features corresponding to the confusion, distraction, enjoyment, neutrality, and fatigue emotions.