Deep Credal Neural Network: Characterization of Imprecision Between Categories

Zuowei Zhang, Zhunga LIU, Arnaud Martin, Kuang Zhou

Keywords: Deep credal neural network, Uncertainty, Imprecision, Belief functions, Meta-category, Classification

TL;DR: We propose a deep credal neural network based on the theory of belief functions to characterize the imprecision in the training and test sets and thereby improve the classification performance.

Abstract: Quantification and reduction of uncertainty in deep learning techniques have received much attention but ignored how to characterize the imprecision caused by such uncertainty. In some tasks, we prefer to obtain an imprecise result rather than being willing or unable to bear the cost of an error. For this purpose, we present a deep credal neural network (DCNN) based on the theory of belief functions, aiming to assign samples that are indistinguishable for specific categories to the union of these, called meta-category. In DCNN, a designed mechanism assigns multiple labels to some training samples to constrain the known loss functions. Once assigned, it indicates that these samples may be in an overlapping region of different categories, or the original label is wrong. Afterward, the training labels are reconstructed and therefore classify the test samples. Once assigned to any meta-category, the prediction of this test sample is imprecise. Experiments based on some remarkable networks have shown that DCNN can not only improve accuracy but also reasonably characterize imprecision both in the training and test sets.