Coherent optical neural networks and the generalization characteristic

Coherent optical neural-network systems are presented by which we are to realize (1) control of optical neural-network behavior by carrier frequency modulation and (2) frequency-domain multiplexing as a new degree of neural parallelism. In the coherent optical neural network, lightwave carries amplitude, phase, and frequency signals which are processed through optical neural connections consisting of transparency, delay, and optical nonlinear neurons. The neural learning process is realized by adjusting the delay time and the transpar-ency of the optical connections by regarding the carrier frequency as a learning parameter. A simulation experiment demonstrates that the proposed system learns desired output signals depending on the optical carrier frequencies, which leads to the above advantages. It is also found that the generalization characteristics depends on built-in input delays. The generalization quality is discussed from the viewpoint of function approximation or synthesis using sinusoidal basis functions.