Table 1 Numerical values when working the Japanese vowels dataset. The first column shows the noise scale (see Equation (8)), the next four columns give, respectively, the accuracy of classification by a noiseless RNN (trained as a stochastic RNN), then the lowest, the largest and the average accuracy given by the (stochastic) RNN over 10 simulations. The following four columns show the same numbers after training the RNN with 10% of mislabelled data. The last column shows the ratio of the average accuracy in the robustness test (after training with some corrupted labels) over the normal average accuracy (after training with the correct labels)

From: Path classification by stochastic linear recurrent neural networks