Adaptive input field neural network-that can recognize rotated and/or shifted character

Presents an adaptive input field neural network which is designed for the classification problem and can recognize rotated and/or shifted characters. It consists of an adaptive input field and an ordinal layered neural network. The input field contains input layer cells which are arranged according to a specific spatial constraint. This model can deal with a rotated and/or shifted input pattern. When such a pattern is presented, the input field is rotated and/or shifted to compensate for initial positional variance. Cells in the input field adjust their positions to reduce the error in the output layer. By utilizing delta values in the input layer, input cell movement direction can be decided. At the same time, the constraint between input cells is introduced to maintain a rectangular shape for the input field grid. Experiments were carried out on the recognition of five alphabetic characters. When an `A´ was rotated 30° and presented to the input field, the input field rotated about 24 degrees against the initial rotation. The network was also effective in recognizing shifted patterns. An `A´ that was shifted diagonally by 20% was also presented, and the input field shifted to the almost-diagonally normal position