Generalization of the VQ-VAE to other data sets tested on digits. Left: A test image from digits 0–4 with different noise levels (left column) is processed by a VQ-VAE network trained on digits 0–4 (in sample), on digits 5–9 (out of sample), or on fashion item classes 0–4 (out of distribution). Right: Noisy images as well as their image reconstructions from the three types of VQ-VAEs are tested for classification accuracy using an MNIST classifier for digits 0–9. The curves are averaged over 10 runs on 10,000 test images, as well as different combinations of training and test set, see main text.