Classification Model Considering Misclassification Based on Distributional Distance Between Labels

Abstract

In this study, we propose deep learning model that not only has a high accuracy, but also a small distance between training and prediction labels for unknown data in the neighbor of the training data. We define a knowledge distribution as probability distribution for each labels. The relationship between the knowledge distributions corresponds to the distance between the labels. In the proposed method, we firstly train a model to minimize distributional distance between internal features and knowledge distribution, then train the model to classify labels correctry. We visually confirmed that the proposed method can predict labels with small distance between labels for unknown data near the training data. In addition, we compared the proposed method with a existing method that uses a typical convolutional neural network. Experiments on MNIST show that the proposed method achieved the same or better accuracy than existing methods. The proposed method also achieved mean absolute error of 0.32, while the existing method achieved 0.374.