Incremental training of Restricted Boltzmann Machines using information driven saccades

In the context of developmental robotics, a robot has to cope with complex sensorimotor spaces by reducing their dimensionality. In the case of sensor space reduction, classical approaches for pattern recognition use either hardcoded feature detection or supervised learning. We believe supervised learning and hard-coded feature extraction must be extended with unsupervised learning of feature representations. In this paper, we present an approach to learn representations using space-variant images and saccades. The saccades are driven by a measure of quantity of information in the visual scene, emerging from the activations of Restricted Boltzmann Machines (RBMs). The RBM, a generative model, is trained incrementally on locations where the system saccades. Our approach is implemented using real data captured by a NAO robot in indoor conditions.