PURPOSE: To simplify the algorithm of learning to constitute computing elements of optical elements by giving a multilayered structure and preventing an inverse propagation signal for learning from including differentiated values.

CONSTITUTION: No differentiated values are included in the inverse propagation signal for learning of n optical neural net having the multilayered structure. A polarizing beam splitter 11, a light receiving means 12 having the threshold processing function, a difference calculating means 13, a light emitting means 14 and a means 15 which receives the signal from the light receiving means 12 to modulate the light are provided as computing elements of this operation, and further, a second polarizing beam splitter 17, a light receiving means 18 having the threshold processing function, a means 19 which receives the signal of the light receiving means 18 to modulate the light, and a light emitting means 20 are provided. Thus, the algorithm of learning is simplified, and optical parts are used as computing elements to perform the learning operation.