Aprendizaje multitarea mediante arquitecturas neuronales basadas en subredes privadas

Abstract

Human learning frequently involves learning several tasks simultaneously; in particular, humans compare and contrast similar tasks for solving a problem. Nevertheless, most approaches to machine learning focus on the learning of a single isolated task, Single Task Learning (STL). Most of them can be formulated from learning several tasks related to the main task at the same time while using a shared representation, Multitask Learning (MTL). This type of learning improves generalization performance for a main task by using the information contained in other related tasks. In this article, we examine distinct schemes used in MTL, propose an new network architecture and test each scheme in two different problems. The proposed scheme makes use of private subnetworks to improve the performance of MTL.