Scaling up a Boltzmann machine model of hippocampus with visual features for mobile robots

Previous papers [4], [5] have described a detailed mapping between biological hippocampal navigation and a temporal restricted Boltzmann machine [20] with unitary coherent particle filtering. These models have focused on the biological structures and used simplified microworlds in implemented examples. As a first step in scaling the model up towards practical bio-inspired robotic navigation, we present new results with the model applied to real world visual data, though still limited by a discretized configuration space. To extract useful features from visual input we apply the SURF transform followed by a new lamellae-based winner-take-all Dentate Gyrus. This new visual processing stream allows the navigation system to function without the need for a simplifying data assumption of the previous models, and brings the hippocampal model closer to being a practical robotic navigation system.