Experiments of learning in optical perceptron-like and multilayer neural networks

An optical multilayer network with backpropagation learning capability is constructed and tested. The modifiable connection weights and the unit plane are realized by using a photorefractive crystal and a microchannel spatial light modulator, respectively. First, in a simple optical perceptron-like network, the experiment of learning is conducted. The learning rate is optimally determined by setting the exposure time of the hologram based on the temporal characteristics of the photorefractive crystal. Next, the experiment is extended to the optical three-layer network. The optimal error signal for backpropagation learning is successfully generated. By incorporating the optical error signal into the network, the experiment of learning is performed. Due mainly to optical losses of the system related to the holograms and the unit planes, the projected performance is not fully realized. The preliminary experiment is individually conducted by partly disconnecting the optical interconnections between the optical elements. However, the key performances to a full optical realization of backpropagation learning are obtained.<<ETX>>