Abstract

We report an experimental two-layer optical tiling network,¹ the hidden layer of which grows during training. The hidden neurons are added like tiles only when they are needed. The first layer of interconnections is implemented by a dynamically refreshed photorefractive (BaTiO₃) volume hologram. The hidden layer starts with a single unit, which is trained with the perceptron algorithm to classify the 2-D training inputs as well as possible. Then a second hidden neuron is added and trained to correct as well as possible classification errors that the first hidden neuron makes. This procedure is repeated until a satisfactory hidden representation of the inputs is obtained. The second layer, implemented with an optical matrix-vector multiplier, is then trained to solve the problem based on the hidden representation. A key feature of this system is the use of dynamic copying² to control the weight decay in the photorefractive crystal as new holograms are formed.

© 1992 Optical Society of America