Whitehill, Jacob Richard
Automatic estimation of student engagement can help computer-based learning systems adapt to individual learners. Linear models trained on Gabor features established cutting-edge yet sub-human accuracy on this task, while Convolutional neural networks (CNNs) heavily overfit to the dataset's few subjects. We found that transfer learning enabled linear ridge regression to leverage CNN features learned for image recognition and face re-identification tasks. Our best model achieved a four-fold cross-validated correlation of r=0.581, significantly outperforming the state-of-the-art r=0.522. Our information strength metric correlated with model accuracy (FaceNet, r=0.755; ImageNet, r=0.077), inviting further study of feature utility prediction.
Worcester Polytechnic Institute
Major Qualifying Project
All authors have granted to WPI a nonexclusive royalty-free license to distribute copies of the work, subject to other agreements. Copyright is held by the author or authors, with all rights reserved, unless otherwise noted.