A cortex-inspired episodic memory toward interactive 3D robotic vision

This paper shows the advantage of using a cortex-inspired episodic memory model in a robotic vision-system. The robot can interact, learn, and recall 3D objects in real-time. The model forms sparse distributed memory traces of spatiotemporal episodes. These episodes consist of sequences of sensorimotor patterns. These patterns represent the visual scenes of 3D objects and the robot states when encountering the objects. The results show: 1) Dynamic recall, when the model is prompted with the initial items of the learned episode. 2) Recognition, by recalling the most similar stored objects when encountering new objects. 3) Sensorimotor learning, by generating the missing information when encountering either similar visual input or similar robot's states. The model learns by measuring the degree of similarity between the current input pattern on each time slice and the expected input given the preceding time slice (G). Then adding an amount of noise, inversely proportional to G, to the process of choosing the Internal Representation of the model.