Rank consistent ordinal regression for neural networks with application to age estimation

In many real-world prediction tasks, class labels include information aboutthe relative ordering between labels, which is not captured by commonly-usedloss functions such as multi-category cross-entropy. Recently, the deeplearning community adopted ordinal regression frameworks to take such orderinginformation into account. Neural networks were equipped with ordinal regressioncapabilities by transforming ordinal targets into binary classificationsubtasks. However, this method suffers from inconsistencies among the differentbinary classifiers. To resolve these inconsistencies, we propose the COnsistentRAnk Logits (CORAL) framework with strong theoretical guarantees forrank-monotonicity and consistent confidence scores. Moreover, the proposedmethod is architecture-agnostic and can extend arbitrary state-of-the-art deepneural network classifiers for ordinal regression tasks. The empiricalevaluation of the proposed rank-consistent method on a range of face-imagedatasets for age prediction shows a substantial reduction of the predictionerror compared to the reference ordinal regression network.