Fig. 3. Accuracy of the networks on MNIST and fashion-MNIST datasets. The test accuracy during a single experiment on (A) MNIST and (B) fashion MNIST was plotted over the training iterations for male and hermaphrodite models. The accuracy converges to >99.5% on MNIST for both models in (A), and to >92.6% on fashion MNIST for both models in (B). A few examples are given for the MNIST digit images that are classified correctly (C) and incorrectly (E), where the digits in (E) are difficult to recognize even for a human. The same number of examples are given also for the fashion MNIST images that are classified correctly (D) and incorrectly (F), where a human can guess most of the items in (E).
© Exp Neurobiol
