Abstract

Neural networks, characterized as a large number of highly interconnected simple processors, can be trained by varying the strength (weight) of the interconnections (synapses) between the simple processors (neurons). Several holographic optical systems have physically demonstrated this capability previously.^[1][2][3][4] Since neural networks are trained by example rather than programmed with specific rules, they are likely to be able to generalize, or recognize patterns that do not exactly match those used for training. Such generalization is important in real world pattern- recognition problems where the size, orientation, position and background cannot be determined in advance.

© 1991 Optical Society of America