Meta Descent Learning for Class Imbalanced Age Estimation

Weiwei Cai, Xin Dong, Hao Liu

Abstract: In this paper, we propose a meta descent learning method (MDL) for class imbalanced age estimation while preserving the relative ordinal information. The class imbalanced problem causes head classes with enough samples dominant the gradient descent process, thus suppressing the performance of tail classes. Due to its superiority for quick adaptive descent optimization, we utilize meta learning to adjust the learning gradient iteratively for balancing training. Specifically, we reweight the sample loss using meta descent learning to prevent the dominance of head classes in feature space. Furthermore, we propose an order-consistent loss to explicitly constrain the ordinal information on the output logits, helping the exploit of semantic correlation in aging images. Experimental results on several datasets including Morph II and ChaLearn LAP demonstrate the effectiveness of our method.