Adaptive Space Reconstruction on Hidden Layer and Knowledge Transfer based on Hidden-level Generalization in Layered Neural Networks

Humans can learn its action effectively in the real world by taking a similar action in a similar situation. In order to realize such abilities in a robot with a neural network, not only generalization on the input signal space, but also the generalization based on the similarity on the reconstructed space on the hidden layer would play an important role. In this paper, to explain the acquisition of the useful hidden representation for the generalization, a hypothesis is set up at first that when two training patterns are closer, the corresponding hidden patterns are likely to become closer through learning. In the first simulation, the hypothesis is supported by the learning of random input-output pattern sets. In the other simulations where simple visual sensor signals were the input of the network, it is shown that the hidden layer represents global information adaptively while keeping the information given by the initial connection weights from the input layer to the hidden layer. Finally, the reason why a neural network with visual inputs becomes to represent the global information in the hidden layer through the reinforcement learning of the task that a robot reaches a target caught on its visual sensor, is considered.