A Hybrid Multi-Layer Architecture for Autonomous Vehicles Utilising a Hierarchical Perception-Action Dreams Simulation Mechanism
Published at EU-Cognition Meeting, Zurich, 23-24th November 2017
Deep neural architectures have been proposed as a solution for Autonomous Vehicles. However, deep learners require very large sets of training data such that meeting current requirements of autonomous vehicle safety is intrinsically hard to achieve. We here propose a hybrid multi-layer architecture, featuring a biologically inspired separation between the tasks of action priming and action selection, that extends the principle of hierarchical Perception-Action learning via a dream simulation mechanism to greatly extend the utility of training data during learning.