Interactive Machine Learning and Explainability in Mobile Classification of Forest-Aesthetics

Simon Flutura, Andreas Seiderer, Tobias Huber, Katharina Weitz, Ilhan Aslan, Ruben Schlagowski, Elisabeth André, Joachim Rathmann

Abstract: This paper presents an application that classifies forest's aesthetics using interactive machine learning on mobile devices. Transfer learning is used to be able to build upon deep ANNs (MobileNet) using the limited resources available on smart-phones. We trained and evaluated a model using our application based on a data-set that is plausible to be created by a single user. In order to increase the comprehensibility of our model we explore the potential of incorporating explainable Artificial Intelligence (XAI) into our mobile application. To this end we use deep Taylor decomposition to generate saliency maps that highlight areas of the input that were relevant for the decision of the ANN and conducted a user study to evaluate the usefulness of this approach for end-users.