Adaptive Space Reconstruction and Generalization on Hidden Layer in Neural Networks with Local Inputs

Our living creatures represent global information in their brain by integrating local sensory signals such as visual sensory signals. In this paper, the state of hidden layer in a layered neural network with local inputs after learning was observed for some cases. Some characters became clear as follows. (1)If the training signal changes gradually in space, the hidden layer becomes to represent the spatial information. (2)This tendency is stronger in the higher hidden layer. (3)If there are redundant hidden neurons, they represent the global information totally, while each of them keeps the initial uctuation due to the initial connection weights. (4)If there is no correlation between the training signal of two input region, the learning of one region becomes not to in uence to the learning of the other region. (5)However, the hidden neurons does not become to represent the information for only one region. From these results, the reason why the hidden layer becomes to represent spatial information by reinforcement learning [1] can be thought as follows. The state evaluation value changes gradually according to the time to the goal, while motion should change gradually for the states with the same evaluation value.