Auxiliary Learning for Hypercomplex Neural Networks

The capability of hypercomplex neural networks to model complex spatiotemporal phenomena is strictly linked to the compact formalization of data obtained by using quaternions. In this contribution, we explore the possibility to optimize the learning of hypercomplex neural networks based on the Levenberg-Marquardt method, in terms of reducing the number of epochs needed to complete learning. An approach based on instantiating parallel hypercomplex networks devoted to predict the trend of weights during the learning phase is outlined. The designed neural architecture is then tested on a hyperchaotic dataset showing the effectiveness of the approach