Constraints-based explanation by visual feature learning

抄録

The paper develops a semi-supervised algorithm with user feedback for feature level explanation generation. Letting human to label data usually will improve learning accuracy. At the same time, if we presume user feedback as background knowledge, there also exists potential to enhance a model's interpretability. However, the acquisition of labeled data often requires a skilled expert. In this paper, we convert the features extracted by deep learning into natural images for human understanding. Then we design a user interface where users can drag and drop image segments on the clustering result. We also propose a method to improve constraints for better clustering. In our experiments, two classes from the large-scale dataset, we explore the benefit of our semi-supervised approach over the image features given by CNN models. This semi-supervised approach not only improves classification accuracy but also helps downstream applications generate explanation.