The Sock Sorter

Sort thirty of your own things into two piles, write the rules a computer would need to tell them apart, then build three objects on purpose that break your own rules.

Easy · 1 hour

Introduction

Every time a phone picks a face out of a photo or an app decides a message is spam, something learned the difference from examples. Not from being told the answer, from being shown piles of things.

You are going to do that same job by hand. Two piles, thirty objects, and a set of rules in your own handwriting. Then you are going to try to beat your own rules, which is where professional AI teams spend most of their time.

The Why

A computer cannot see a sock. It can only measure things about a sock: how long it is, how many colors it has, whether it has a heel. The measurements you choose are called features, and picking them is most of the work. The objects you invent at the end are called edge cases, and they are where real systems fail. Hunting them on purpose is how you learn that a confident answer and a correct answer are not the same thing.

Step-by-Step Instructions

1
Pick two groups you can tell apart at a glance: socks and gloves, forks and spoons, coins and buttons. Gather about fifteen of each and split them into two piles.
2
Photograph or draw each pile so you have a record. This set is the only thing your rules are allowed to learn from, so it is worth keeping.
3
Without touching a computer, list every measurement you could take of an object: length in centimeters, number of colors, number of holes, whether it bends, whether it is shiny. These are your features.
4
Now write your rules, using only those measurements. For example: if it is longer than 12 cm and has no holes, it is a sock. Keep each rule short enough to read out loud in one breath.
5
Test your rules on all thirty objects, one at a time. Mark every object your rules get wrong and count them. That count is your error rate.
6
Change one rule, then test all thirty again. Write down both scores so you can see whether the change actually helped.
7
Now attack your own work. Find or build three objects your rules will get wrong on purpose: a very short sock, a glove with the fingers tucked in, something that is somehow both. Test them, and write down what your rules said and why they were fooled.

Rule Test Log

One row per test round. Run your rules over all thirty objects, count what they got wrong, then change one rule and run them again.

Round	Rule you changed	Objects tested	Got it right	Got it wrong
Row 1, Round	Row 1, Rule you changed	Row 1, Objects tested	Row 1, Got it right	Row 1, Got it wrong
Row 2, Round	Row 2, Rule you changed	Row 2, Objects tested	Row 2, Got it right	Row 2, Got it wrong
Row 3, Round	Row 3, Rule you changed	Row 3, Objects tested	Row 3, Got it right	Row 3, Got it wrong
Row 4, Round	Row 4, Rule you changed	Row 4, Objects tested	Row 4, Got it right	Row 4, Got it wrong

If a round makes the score worse, keep the row anyway. A change that hurt is evidence too, and deleting it is how people end up fooling themselves.