The optical implementation of a neural network consists of two basic components: a 2-D array of neurons and interconnections. Each neuron is a nonlinear processing element that, in its simplest form, produces an output which is the thresholded version of the input. Liquid crystal spatial light modulators, optoelectronic integrated circuits (OEIC's), either hybrid, such as liquid crystal on silicon, Si-PLZT, and flip-chip devices, or monolithic integration in 111-V compounds, are examples of such a solution. In order for these devices to be used as neurons in a practical experiment, they must contain a large number of neurons (104/cm2 - 106/cm2) and exhibit high gain. This puts a stringent requirement on the electrical power dissipation. Thus, these devices have to be operated at low enough current levels so that the power dissipation on the chip does not exceed the heat-sinking capability, and yet the current levels need to be large enough to be able to produce high gain. This means sensitive input devices are a must. To achieve these goals, the speed requirement of the devices must be relaxed as the operation of neural network does not have to be too fast.