Learning models: Explainability for Black-Box Neural Network Learners in the Education Domain
Event details
| Date | 26.07.2021 |
| Hour | 16:00 › 18:00 |
| Speaker | Vinitra Swamy |
| Category | Conferences - Seminars |
EDIC candidacy exam
exam president: Prof. Pierre Dillenbourg
thesis advisor: Prof. Tanja Käser
thesis coadvisor: Pror. Martin Jaggi
co-examiner: Prof. Antoine Bosselut
Abstract
Digital learning environments are commonplace in the modern classroom. The rise of educational technology has been subsequently mirrored by the rise of applied machine learning research in areas like student learner modeling, autograding, dropout prediction, and curriculum design. Although the amount of research in this sphere has grown significantly, the adoption of neural network models in learning platforms has not yet become ubiquitous. Critics of machine learning in education are concerned about the interpretability of black box models and the privacy of student data. Other educators do not see how large scale models can have an impact on student performance prediction in their small, ongoing, or first-time course.
In this doctoral candidacy proposal, we present three papers aimed at overcoming the gap between real-world educational data challenges and intelligent predictors of student performance. The first paper addresses the problem of unknown student outcomes in ongoing courses using transfer learning to demonstrate that knowledge can be shared across MOOCs for student dropout prediction. The second paper addresses the small classroom size problem with an active-learning approach, showing that highly performant student affect detectors can be trained using a minimal set of data points. The third research paper is a landmark work in the explainable AI field, focusing on traditionally interpretable local models (LIME) to explain black box model behavior. Although LIME has enjoyed great popularity in the ML research community, there has not been much work in neural network explainability for education. We build upon these works to propose a research agenda overcoming practical adoption concerns in the machine learning for education field through transfer learning, active learning, and interpretability.
Background papers
exam president: Prof. Pierre Dillenbourg
thesis advisor: Prof. Tanja Käser
thesis coadvisor: Pror. Martin Jaggi
co-examiner: Prof. Antoine Bosselut
Abstract
Digital learning environments are commonplace in the modern classroom. The rise of educational technology has been subsequently mirrored by the rise of applied machine learning research in areas like student learner modeling, autograding, dropout prediction, and curriculum design. Although the amount of research in this sphere has grown significantly, the adoption of neural network models in learning platforms has not yet become ubiquitous. Critics of machine learning in education are concerned about the interpretability of black box models and the privacy of student data. Other educators do not see how large scale models can have an impact on student performance prediction in their small, ongoing, or first-time course.
In this doctoral candidacy proposal, we present three papers aimed at overcoming the gap between real-world educational data challenges and intelligent predictors of student performance. The first paper addresses the problem of unknown student outcomes in ongoing courses using transfer learning to demonstrate that knowledge can be shared across MOOCs for student dropout prediction. The second paper addresses the small classroom size problem with an active-learning approach, showing that highly performant student affect detectors can be trained using a minimal set of data points. The third research paper is a landmark work in the explainable AI field, focusing on traditionally interpretable local models (LIME) to explain black box model behavior. Although LIME has enjoyed great popularity in the ML research community, there has not been much work in neural network explainability for education. We build upon these works to propose a research agenda overcoming practical adoption concerns in the machine learning for education field through transfer learning, active learning, and interpretability.
Background papers
- Transfer Learning using Representation Learning in Massive Open Online Courses
- Authors: Mucong Ding, Yanbang Wang, Erik Hemberg, Una-May O'Reilly
- Link: https://arxiv.org/pdf/1812.05043.pdf
- LIME: "Why Should I Trust You?": Explaining the Predictions of Any Classifier
- Authors: Marco Tulio Ribeiro, Sameer Singh, Carlos Guestrin
- Link: https://arxiv.org/pdf/1602.04938.pdf
- Active Learning for Student Affect Detection
- Authors: TY Yang, RS Baker, C Studer, N Heffernan, AS Lan
- Link: https://www.research-collection.ethz.ch/bitstream/handle/20.500.11850/461351/19EDM-al.pdf?sequence=1&isAllowed=y
Practical information
- General public
- Free