Visual Explanations for DNNs with Contextual Importance

Sule Anjomshoae, Lili Jiang, Kary Främling

Abstract: Autonomous agents and robots with vision capabilities powered by machine learning algorithms such as Deep Neural Networks (DNNs) are taking place in many industrial environments. While DNNs have improved the accuracy in many prediction tasks, it is shown that even modest disturbances in their input produce erroneous results. Such errors have to be detected and dealt with for making the deployment of DNNs secure in real-world applications. Several explanation methods have been proposed to understand the inner workings of these models. In this paper, we present how Contextual Importance (CI) can make DNN results more explainable in an image classification task without peeking inside the network. We produce explanations for individual classifications by perturbing an input image through over-segmentation and evaluating the effect on a prediction score. Then the output highlights the most contributing segments for a prediction. Results are compared with two explanation methods, namely mask perturbation and LIME. The results for the MNIST hand-written digit dataset produced by the three methods show that CI provides better visual explainability.