Shared Roots: Regularizing Deep Neural Networks through Multitask Learning

In this paper, we propose to regularize deep neural nets with a new type of multitask learning where the auxiliary task is formed by agglomerating classes into super-classes. As such, it is possible to jointly train the network on the class-based classification problem AND super-class based classification problem. We study this in settings where the training set is small and show that , concurrently with a regularization scheme of randomly reinitializing weights in deeper layers, this leads to competitive results on the ImageNet and Caltech-256 datasets and state-of-the-art results on CIFAR-100.