Recursive Division Explainability as a Factor of CNN Quality

In this paper, we present the result of the research related to the possibility to measure numerically the explainability of already trained CNN model using methods based on the perturbations of input image. The level of explainability is measured as the average IOU over the binarized classification explanations provided by RISE, Grad-CAM, RD methods, and ground-truth labels of the dataset. The evaluation of explainability is performed for five networks having different architectures and accuracies and trained to solve the same cat/dog classification problem for Oxford-IIIT Pet Dataset with 37 classes. It has been shown that CNN with high accuracy does not necessarily have good IOU explainability. The decision about which model is better is split between different methods. The question about measuring the explainability becomes more acute when ground-truth labels are missing for the dataset, so qualitative evaluations with IOU are not possible. For this case the proposed RDE value may be used as some indicator of the explainability of the model. It is also shown that the performance of RD is better compared to other methods.