Parallel distributed processing model with local space-invariant interconnections and its optical architecture

This paper proposes a parallel distributed processing model with local space-invariant interconnections, which is more readily implemented by optics and is able to classify patterns correctly, even if they have been shifted or distorted. Error backpropagation is used as a training algorithm. Computer simulation results presented indicate that the processing is effective and the network can deal with the shifted or distorted patterns. Moreover, the optical implementation architecture using matched filters for the model is discussed.