Classifying Handwritten Digits without Seeing Them

Digital images of handwritten digits are high dimensional and vary with writing style. This work presents a method to perform classification on a handwritten digits database, MNIST, and enable visualization in low dimensional space. To address the handwritten digit variation issue, the edges of each digit in an image are first highlighted by gradient feature extraction. Then, the curse of high dimension is broken by t-SNE algorithm, which constructs a certain “lens” so that one can visualize MNIST on two or three coordinates. The “lens” also helps trace from low dimension back to high dimension in which clustering is applied to assigned level sets and form a more explicit visible structure among all data points. The last process is done by Mapper algorithm