Figure 1. The MCF–based in-fiber neural network learns with external feedback and learning algorithm.

During learning or adapting (if required), the outputs are sampled by detectors and delivered as electronic signals to a PC or a controller, which computes the necessary change in amplification pattern and replaces synaptic weights (top). After learning is optimized, the device can work “on-line” without further change of weights, thus no longer requiring feedback or a learning algorithm (bottom).