Quasi-Unimodal Distributions for Ordinal Classification

Ordinal classification tasks are present in a large number of different domains. However, common losses for deep neural networks, such as cross-entropy, do not properly weight the relative ordering between classes. For that reason, many have been proposed literature, which model output probabilities following unimodal distribution. This manuscript reviews these on three datasets and suggests potential improvement focuses constraint neighborhood around true class, allowing more flexible distribution, aptly called quasi-unimodal loss. this purpose, two constraints proposed: A first concerns order top-three probabilities, second ensures remaining higher than top three. Therefore, gradient descent improving decision boundary class detriment to distant The loss is found be competitive several cases.