Optical Realizations of Perceptron-like Neural Networks
- 168 Downloads
The perceptron learning and network algorithm is one of the most simple and popular structures in optical neural network relizations. Although it allows only for processing of linear decision problems because it is based on a simple outer product summation of the inputs followed by a nonlinear thresholding operation (see section 2.2), it is the basis element for all further multilayer nets and easy to implement optically using different interconnection techniques. An important feature of perceptron algorithms is that they allow both additive and subtractive changes to the weights (see section 4.2.4) . This bipolar weight realization allows for excitatory as well as for inhibitory interconnections and therefore for bipolar weight changes. It is the realization of that feature that is required in most of the multiple layer networks such as backpropagation nets and will play an important role for all-optical realizations
KeywordsOptical Realization Spatial Light Modulator Beam Coupling Photorefractive Crystal Input Plane
Unable to display preview. Download preview PDF.
- 1.B.K. Jenkins and A.R. Tanguay, Photonic implementations of neural networks, in Neural networks for signal processing, B. Kosko, ed., pp. 287 - 382, Prentice Hall, Englewood Cliffs, NJ (1992).Google Scholar