Coherent optical neural networks: generalization ability in frequency domain

Abstract

Frequency-domain generalization ability in coherent optical neural networks is analyzed. The coherent optical neural network system consists of an optical complex-valued neural network, a phase reference path, and coherent detectors with 90-degree optical hybrids for self-homodyne detection. The learning process is realized by adjusting delay time and transparency of neural connections in the complex-valued neural network. Information geometry in the learning process is discussed for obtaining a parameter region where a reasonable generalization is realized in frequency space. It is found that there are error-function minima appear periodically both in delay-time domain and input-signal-frequency domain and, hence, that initial connection delay should be within a certain range for a successful learning. Experiments demonstrate that a stable learning and a reasonable generalization in the frequency domain are realized in a parameter range suggested by the theory.