Automatic recognition of handwritten digits is an important problem, which can be found in many practical applications. In this chapter, we will implement some feed-forward networks to address this problem.
To train and test the implemented models we use the MNIST database of handwritten digits.
The MNIST dataset is made of a training set of 60,000 examples, plus a test set of 10000 examples. An example of the data, as it is stored in the files of the examples, is shown in the following figure:
The source images were originally in black and white, but later, to normalize them to the size of 20 × 20 pixels, intermediate brightness levels have been introduced, due to the effect of the anti-aliasing filter for resizing. Subsequently, the images were focused in the center of mass of the pixels, in an area of 28×28 pixels...