Investigation into effects of device variation on performance of optoelectronic neural networks

Free space optical links, using LEDs and photodiodes, have been studied as a means of communication between neural layers in an analogue multilayered perceptron neural network architecture, trained using the backpropagation learning algorithm. Simulation results on network learning and generalisation show that such networks are tolerant to as much as 75% variation in either the LED or photodiode responses.