Poster
Authors: Martin Ennemoser, Peter Ruch, Holger Stitz, Hendrik Strobelt, Marc Streit
Abstract: In order to monitor the learning process and track model quality, training of a neural network on a classification task is usually accompanied by accuracy and loss curves and the performance of the final model is summarized using a confusion matrix. However, showing the final result only completely disregards the change (flow) of the model confusion across epochs of the learning process. We propose ConfusionFlow, a generalization of the confusion matrix concept over time that enables the user to uncover the learning dynamics of the neural network model. As a first step towards a more informed design process for network architectures and selection of an optimization procedure and its hyperparameters, ConfusionFlow allows for interactive comparative, exploration of model confusion over the networks learning process.
