Coarse-coding applied to HONNs for PSRI object recognition

Summary form only given, as follows. It is noted that a higher-order neural network (HONN) can be easily designed for position, scale, and rotation invariant (PSRI) object recognition. Invariances are built directly into the architecture of a HONN and do not need to be learned. Fewer training passes and a smaller training set are therefore required to learn to distinguish between objects. The size of the input field is limited, however, because of the memory required for the large number of interconnections in a fully connected HONN. By using coarse coding, the input field size can be increased to allow the larger input scenes required for practical object recognition problems. Using coarse coding, the size of the input field was increased to 127×127 pixels and a third-order neural network was trained to distinguish between a `T´ and a `C´ independent of distortions in translation, in-plane rotation, or scale up to a factor of four. In addition, the same network achieves invariance between a number of other objects, including distinguishing between an F18 aircraft and a Space Shuttle Orbiter. In each case, the network is trained on just one view of each object and learns to distinguish between the two objects in fewer than ten passes through the training set