Iterative and interactive feedback model for the personalization of deep learning

Abstract

We proposed the novel feedback model for Convolutional Neural Networks (CNNs) in doing image classification that requires the model explainer, and has the purpose of allowing user to personalize the prediction made by CNNs through the iterative and interactive feedbacks submitted to the machine. The feedback model is practically the response matrix inserted in between two consecutive layers of CNNs, which will continuously process the feedback obtained and manifest it through the updated prediction behavior. The result showed that the method successfully fulfills each of the user’s objective in personalizing the prediction of the network and furthermore improve the explanation towards that image. Moreover, upon deploying the result of those iterations to be tested on several different images, we observed that none of those images have the explanation worsened. Some of the explanations were even improved after those iterations have been done.