Semantic Spatiotemporal Memory toward 3D Robotic Vision

Abstract

3D robotic vision is proposed using a neural network model that forms sparse distributed memory traces of spatiotemporal episodes of an object. These episodes are generated by the robot interaction with the environment or by robot's movement around 3D object and its perspective to the objects. The traces are distributed in each cell and synapse that participates in many traces. This sharing of representational substrate enables the model for similarity based generalization and thus semantic memory. The results are provided showing that spatiotemporal patterns map to similar traces, as a first step for robot 3D vision system. The model achieves this property by measuring the degree of similarity between the current input pattern on each frame and the expected input given the preceding frame and then adding an amount of noise, inversely proportional to the degree of similarity, to the process of choosing the internal representation for the current frame and the predictable input given the preceding frame.