Abstract

Neural networks require two types of operations; interconnections, which define how the output of one state affects the input of the next, and non-linear operations, which relate the inputs of a state to its output. Interconnections, which require many signals passing through the same space, are best performed with photons, which do not interact with one another. Non-linear operations require interaction (i. e., cross products) between the various inputs to a state, and are best performed with electrons, which interact strongly through their electrical charge. In a typical neural network architecture, almost all of the computation required is associated with the interconnections, and only a tiny fraction is associated with the non-linear operations (sigmoidal response or thresholding) performed at each state.

© 1991 Optical Society of America