A sensory information processing system using neural networks

In order to carry out actions particular to the goals, a robot processes sensory information, that is, it transforms sensed data to internal representation. In some cases, the robot´s internal representation cannot be determined uniquely from the sensed data. An architecture is proposed for a sensory information processing system that overcomes this ill-posed problem. The system uses an artificial neural network which is trained to transform internal representation to sensory data. Applying an iterative scheme to the network, the unique internal representation can be determined. The scheme compares the network´s output (sensory data) with the sensed data, and by backpropagating the difference through the layers updates an input (internal representation) which could have created the applied output (sensed data) based on the gradient descent method. By predicting the resulting state based on the intention of the system´s own movement, the accuracy and speed of sensory information processing can be improved. Simulation results for three-dimensional object recognition are given