Rank-consistent Ordinal Regression for Neural Networks

In many real-world predictions tasks, class labels include information about the relative ordering between labels, which is not captured by commonly-used loss functions such as multi-category cross-entropy. Recently, ordinal regression frameworks have been adopted by the deep learning community to take such ordering information into account. Using a framework that transforms ordinal targets into binary classification subtasks, neural networks were equipped with ordinal regression capabilities. However, this method suffers from inconsistencies among the different binary classifiers. We hypothesize that addressing the inconsistency issue in these binary classification task-based neural networks improves predictive performance. To test this hypothesis, we propose the COnsistent RAnk Logits (CORAL) framework with strong theoretical guarantees for rank-monotonicity and consistent confidence scores. Moreover, the proposed method is architecture-agnostic and can extend arbitrary state-of-the-art deep neural network classifiers for ordinal regression tasks. The empirical evaluation of the proposed rank-consistent method on a range of face-image datasets for age prediction shows a substantial reduction of the prediction error compared to the reference ordinal regression network.