Learning in Optical Neural Networks

In this paper we will review recent advances in training optical neural networks. We will focus on holographic implementations using photorefractive crystals [1]. The vast majority of learning algorithms in neural networks are based on some form of generalized “Hebbian Learning”. With Hebbian learning the strength of the connection between two neurons is modified in proportion to the product (or possibly some other simple function) of the activation functions of the two neurons. These activation functions are typically the neuron response and error signals. The multiplicative Hebbian rule can be implemented if the hologram that connects two neurons is formed as the interference of two light beams generated by the two neurons. This simple and elegant method for training an individual connection can also form the basis for training large optical networks. There are several issues that need to be addressed however before such networks can be constructed. The following is a partial list of these issues, assuming photorefractives are selected as the synapse medium: 1. Architectures for Multiple Holographic Interconnections with 2-D and 3-D Media. 2. Recording Dynamics and Hologram Dynamic Range. 3. Suitable Devices for Neuron Implementation.