R Deep Learning Essentials - Second Edition

Different data distributions

In previous chapters, we used the MNIST dataset for classification tasks. While this dataset contains handwritten digits, the data is not representative of real-life data. In Chapter 5, Image Classification Using Convolutional Neural Networks, we visualized some of the digits, if you go back and look at these images, it is clear that these images are in a standard format:

• There are all grayscale
• The images are all 28 x 28
• The images all appear to have at border of at least 1 pixel
• The images are all of the same scale, that is, each image takes up most of the image
• There is very little distortion, since the border is black and the foreground is white
• Images are the right way up, that is, we do not have any major ...