Convex Multi-Task Learning with Neural Networks

AbstractMulti-Task Learning aims at improving the learning process by solving different tasks simultaneously. The approaches to Multi-Task Learning can be categorized as feature-learning, regularization-based and combination strategies. Feature-learning approximations are more natural for deep models while regularization-based ones are usually designed for shallow ones, but we can see examples of both for shallow and deep models. However, the combination approach has been tested on shallow models exclusively. Here we propose a Multi-Task combination approach for Neural Networks, describe the training procedure, test it in four different multi-task image datasets and show improvements in the performance over other strategies.KeywordsMulti-task learningDeep learningConvex combination