Extended Data Fig. 5: Effect of input noise on RNN predictions.
From: Predicting ultrafast nonlinear dynamics in fibre optics with a recurrent neural network
The left and right panels shows how ± 10 and ± 20 % relative random multiplicative intensity noise added to the examples shown in Extended Data Fig. 2(a) and (c) affect the RNN predictions (see also Methods). The r.m.s. error computed over 50 test realizations was R = 0.200 and R = 0.271 for the ± 10 and ± 20 % cases, respectively (R = 0.152 for noise-free data). Although we do note a residual shift in the point of maximum compression, the dynamics of the higher-order soliton compression are overall well reproduced even under noisy conditions.
