Figure 5
From: Path classification by stochastic linear recurrent neural networks
Verification of the theoretical PAC-bound (Theorem 3.8) and accuracy of the network for the Japanese vowels dataset. Each row of figures represents the results for different noise scales (as defined in Equation (8)); 1, 1.5 and 2.5, respectively. The middle column shows the evolution of the accuracy of the network when increasing the size of the training dataset. The right column shows how an increasing portion (described as mislabelling frequency) of corrupt labels in the training dataset affects the accuracy on the test set. Solid lines show the accuracy of classification by a noiseless RNN (trained as a stochastic RNN). Dotted lines show the results of simulations of the realised labels by the stochastic RNN