ADAPTIVE INPUT FIELD NEURAL NETWORK

Many neural network models for classification problems have been proposed. However, only a few models, such as Neocognitron, DNN, TDNN and layered neural nets of ADALINE can appropriately handle position variance and image distortion. In these models, an adaptation mechanism is built into the neural network. The author presents an Adaptive Input Field Neural Network (AIFNN); this neural network is for the classification problem, and consists of an adaptive input field and a conventional layered neural network. The AIFNN can recognize shifted or distorted signals. The input field contains input layer cells which are arranged according to specific spatial constraint. When such a distorted image is presented, the input field compensates for initial positional distortion, and adapted input field indicates the position where the categorized signal is located. By utilizing delta values in the input layer, the direction of movement of input cells can be decided to adapt for input signal. At the same time, a constraint between input cells is introduced, to maintain a relative spatial position for the input field mesh. This mechanism can also be applied for input signal offset and contrast adjustment. Since the AIFNN is based on the delta value, the network can be any model, such as TDNN or recurrent neural network. Experiments are carried out on the recognition of 26 alphabetical characters (printed by a printer). When an "A" which is slightly moved and 15 degrees rotated is presented to the input field, the input field rotates about 11 degrees against the initial rotation. When a character from a different font set than the learned one is presented to the AIFNN, it recognizes that character and compensates for initial distortion.