Figure 5
From: Photonic machine learning implementation for signal recovery in optical communications
Simulated RC performance of a reservoir with short time delays. The number of samples used per bit representation (j) is disengaged from the number of virtual nodes (k) per time delay τ that are used for training (Methods). The input sequence tested here is for the case of the short-reach transmission system and for \({z^{\prime} }_{1}\) = 50 km (Supplementary Figs 1, 3). Training is performed on a 9-bit reservoir response, as presented in Fig. 3a. Here we map the data recovery performance in BER terms of the test bit stream, by changing two parameters of the reservoir operation: the frequency detuning Δf between the injection and the reservoir laser and the feedback ratio k f (Methods). There is a significant dependence of the BER performance on the transient states used for training and the samples used for the bit representation. For k < 32, RC processing fails to provide an error-free data recovery, even when we incorporate more samples of the input stream. Still, RC provides a significant BER improvement as long as we are using a considerable number of transient states in the training. For comparison, when we apply 9-bit training without the use of the reservoir, BER of the recovered bit stream is higher than 0.02 for all (j, k) representations. BER performance below 10−4 is indicated with the white-shaded colouring.
