FST Graduate Research Topic

Research Topic: Explainable AI (xAI)
Suggested By: Vive Kumar
Last Update: May 17th, 2023

Modern AI techniques that drive cognification are not perfect. Mostly, lack of quality data causes failure. Alternatively, fail can come from issues around the quality and richness of models. As the world of education becomes more evidence-based and more data-centric, the application of cognification is expected to yield more trustworthy outcomes.     Recent advances in explainable AI (xAI) could unveil the inner workings of the underlying AI techniques. Teachers and students can already access such open AI models - for instance, students could revise their work based on xAI feedback prior to submitting assignments and teachers could subject student responses to automatic evaluation of targeted rubrics. xAI techniques could potentially learn from their own explanations, thus leading to the possibility of an autonomous self-improving system. However, at this time of its development, xAI will find an acceptable level of application when tuned with a human-in-the-loop approach.   As an example of xAI, one can consider Automated Essay Scoring (AES) (Kumar & Boulanger, 2020a; 2020b). One can predict the rubric scores of English essays by applying deep learning techniques over a vast range of writing features. Based on thorough analyses of the distributions of rubric score predictions and distributions of resolved and human raters’ rubric scores, one can contend that the rubric scoring models closely approximate the performance of human raters. AU researchers, through studies, have revealed that rubric score prediction does not directly depend on a few word count-based written language features (all word count features were pruned). Many intuitive features were found and selected by each rubric with no dominant features.   This project will be of interest to you if you are interested in studying explainable AI techniques such as SHAP and apply the same with real-world data.