Reinforcement learning in a self-organised representation of feature space

We propose a method of integrating top-down and bottom-up attention, presenting a robotic implementation of a biologically in-spired active vision system. Digits are clas-sified with a self-organised map of features and working memories of rewarding features are maintained as neural activity in the same topological arrangement. From these maps, a retinotopic saliency map is generated, while inhibition of return encourages exploration of the scene. 1