Chapter 6
Getting Real

It’s time to make the leap from simple pizza-related examples to the cool stuff: image recognition. We’re about to do something magical. Within a few pages, we’ll have a program that classifies images.

In this chapter and the next, we’ll apply our binary classifier to MNIST, a database of handwritten digits.^[13] Just a few years ago, before ML systems could tackle more complex datasets, AI researchers used MNIST as a benchmark for their algorithms. In this chapter, we’ll join that esteemed crowd.

 

“Not MNIST Again!”

Besides being a common benchmark in the industry, MNIST is a bit like the “Hello, World!” of ML tutorials. In fact, if you had previous exposure to ML, you might roll your eyes at the sight of yet another example based on MNIST. Why couldn’t I find a more original dataset for this chapter?

Truth is, I chose MNIST because it’s common, not in spite of that. While most tutorials crack computer vision with the help of high-level ML libraries, we’re going to discuss each and every line of code—so we’d better make things easier for ourselves and use a simple, well-documented dataset. Also, a well-known dataset makes it easier for you to look up and compare alternative examples on the Internet.

 

This will be our first experience with computer vision, so let’s start small. We’ll begin by recognizing a single MNIST digit, and leave more general character recognition to the next chapter.