Skip to main content

Methodology

Our three XAI techniques were run on two different domains: MOOC, a tabular dataset where student metrics are used to predict if they will complete a online course, and a set of Image Classifications Models all based round ResNet. Each domain had a seperate user study. To explore the methodology of each model and user study, explore the articles here.

📄️ ResNet Animal Classification

For the image track, we implemented a Residual Neural Network (ResNet) architecture, a derivative of the basic convolutional neural network on a range of image classification tasks. This network was trained on ImageNet, an extremely common dataset for image classification, and will be used in conjunction with a dataset for tumor classification such as that of Panigrahi (2021) and Sarta (2020). Coined by He et al. (2015), ResNet addresses the issues of neural networks’ degradation, where training and evaluation losses spike and lose predictability as a model’s depth passes a certain point. The proposed solution consists of identity mapping, whereby layers pass their outputs to a mapping function that performs one of two actions: it may shortcut part of the input to be recombined at a deeper layer (discussed in Hochreiter et al., 1997), or it may utilize a nonlinear function which asymptotically approximated the use of multiple linear layers at once (introduced in Jégou et al., 2012). This development allows deep neural networks to have no worse loss than shallow ones of appropriate depth.