Module-based neural architectures: some effects by the cooperation of static and dynamical neurons

摘要

This paper introduces two models of module-based neural networks employing static and dynamical neurons, and discusses what effects can be produced due to the integration of mapping and relaxation. In the first model, relaxation is carried out in the internal space warped by mappers or mapping networks. This architecture has the possibility of helping the dynamics avoid inadequate minima. In the second model, two (or multiple) dynamical module-networks are cross-coupled via mapping internetworks. This architecture can produce more stable fixed-point attractors in the network dynamics. Based on some effects in the concrete models, the present paper further discusses the possibility of designing neural network models with more advanced functions via the integration of fundamental principles.