Selective attention of high-order neural networks for invariant object recognition

Summary form only given. Selective attention can be used to reduce the number of inputs for a high-order neural network. By selecting an appropriate scanning mechanism, invariance to translation can be developed. Using a high-order neural network, rotation invariance can be achieved by encoding proper constraints on the connections of receptive fields. The authors have implemented a second-order recurrent neural network to recognize pixel based objects at any translation and 90° rotation and have tested the network with the TC problem