A Neural Network for Learning and Recognizing Rotated Patterns

This paper proposes a 3-layered neural net model that can recognize rotated patterns by learning only standard patterns. The weights of connections are learned by the error back-propagation algorithm under some equivalence constraints to realize the rotation invariant nets. Since this model removes the influence of rotation and descriminates patterns parallely and distributedly, it has following merits : its design is easier, the net size is smaller, learning and recognizing time is shorter than the conventional sequential models that have a fixed rotation invariant net followed by an adaptive net.