Spatio-temporal sequence learning of visual place cells for robotic navigation

In this paper, we present a novel biologically-inspired spatio-temporal sequence learning architecture of visual place cells to leverage autonomous navigation. The construction of the place cells originates from the well-known architecture of Hubel and Wiesel to develop simple to complex features in ventral stream of the human brain. To characterize the contribution of each feature towards scene localization, we propose a novel significance analysis based on the activation profiles of features throughout the spatio-temporal domain. The K-iteration Fast Learning Neural Network (KFLANN) is then used as a Short-Term Memory (STM) mechanism to construct our sequence elements. Subsequently, each sequence is built and stored as a Long-Term Memory (LTM) cell via a one-shot learning mechanism. We also propose a novel algorithm for sequence recognition based on the LTM organization. The efficiency and efficacy of the architecture are evaluated with the vision dataset from ImageCLEF 2010 Competition.