While deep neural networks (DNNs) have been shown to outperform humans on many vision tasks, theirintransparent decision making process inhibits wide-spread uptake, especially in high-risk scenarios. The BagNetarchitecture was designed to learn visual features that are easier to explain than the feature representation of otherconvolutional neural networks (CNNs). Previous experiments with BagNet were focused on natural images providingrich texture and color information. In this paper, we investigate the performance and interpretability of BagNet on a dataset of human sketches, i.e., a data set with limited color and no texture information. We also introduce a heatmapinterpretability score (HI score) to quantify model interpretability and present a user study to examine BagNetinterpretability from user perspective. Our results show that BagNet is by far the most interpretable CNN architecture inour experiment setup based on the HI score.
展开▼
