MNIST Dataset

The MNIST dataset (Lecun et al., 1998), which consists of 60,000 gray-scale images of handwritten digits. Each image has an accompanying label in {0, 1,..., 9}, and is stored as a 28 × 28 matrix that takes on values in [0, 255]. We first divide the entries of all the images by 255. Next, since there is no variation in the peripheral pixels of the images (Gallaugher & McNicholas, 2018), which violates model (1), we add an independent perturbation N (0, 0.01) to each element of the image. Finally, we vectorize each image to obtain a vector xi ∈ R784.