Embedding deep networks into visual explanations

In this paper, we propose a novel Explanation Neural Network (XNN) to explain the predictions made by deep network. The XNN works learning nonlinear embedding of high-dimensional activation vector network layer into low-dimensional explanation space while retaining faithfulness i.e., original can be constructed from few concepts extracted our We then visualize such for human learn about high-level that is using make decisions. an algorithm called Sparse Reconstruction Autoencoder (SRAE) space. SRAE aims reconstruct part feature faithfulness. A pull-away term applied bases more orthogonal each other. visualization system introduced understanding features in proposed method CNN models image classification tasks. conducted study, which shows approach outperforms single saliency map baselines, and improves performance on difficult Also, several metrics are evaluate explanations quantitatively without involvement.