Gradient descent decomposition for multi-objective learning

Abstract

Multi-objective learning has been explored in neural network because it adjusts the model capacity providing better generalization properties. It usually requires sophisticated algorithms such as ellipsoidal, sliding-mode, genetic algorithms, among others. This paper proposes an affordable algorithm that decomposes the gradient into two components and it adjusts the weights of the network separately. By doing so multi-objective learning with L 2 norm control is accomplished. © 2011 Springer-Verlag.