The handwritten-digit problem
What you’ll learn
Section titled “What you’ll learn”This is the first lesson of Track 11 (Neural Network Intuition), and it sets up everything that follows. It starts from something you do without effort, reading a sloppy handwritten digit, and shows why that effortless act is almost impossible to write down as a set of rules. Along the way you will see what a digit actually looks like to a computer (a long list of brightness numbers, with no shapes in it anywhere), why every honest attempt at a recognition rule breaks on real handwriting, and the single change in approach that makes neural networks, and most of modern AI, work: stop describing the answer, start showing labeled examples.
Where this fits
Section titled “Where this fits”This is the opener of the track, so there is nothing before it. Its job is to plant the problem and the central reframe before any mechanism arrives. The lesson deliberately stops at the edge of the solution: it names the thing we are after, a function that turns 784 brightness numbers into 10 scores, but does not yet open it up. The next lesson begins to crack that function open, showing it is built from layers of simple units. The rest of the track builds the full mechanism, one careful step at a time, all the way to how such a system actually learns.
Before you start
Section titled “Before you start”Prerequisites: none. You do not need any math background, any coding background, or anything installed. The practice is pen and paper. If you have ever tried to write step-by-step instructions for something you find obvious, you already have the only intuition this lesson leans on.
By the end, you’ll be able to
Section titled “By the end, you’ll be able to”- Explain why recognizing a handwritten digit is effortless for a person but brutally hard to specify as explicit rules
- Describe what a digit image is to a computer (a grid of brightness numbers, often 784 for a 28 by 28 image) and why the same digit produces different numbers each time
- Articulate the shift from writing rules to showing labeled examples, and why it underlies almost all of modern AI
- Recognize why example-trained systems are strong on fuzzy human tasks yet brittle on inputs unlike their training data
Time and difficulty
Section titled “Time and difficulty”- Read time: about 8 minutes
- Practice time: about 15 minutes (a write-a-rule-and-watch-it-break exercise, a sort-the-tasks drill, and flashcards)
- Difficulty: intro