Beginner course · Chapter 3 of 8
How Machines Learn
Train a model yourself. Then find out what it gets right, and what it gets wrong.
This page lists what a learner is able to do at the end of this class. It does not carry the class itself. The lessons, exercises and answers stay in the classroom and in the book.
What you will be able to do
- Train, test and retrain a machine learning model on your own
- Explain what a model actually learns, and what it does not
- Say why varied examples beat more examples of the same kind
- Explain bias in plain language and give a real example of it
- Recognise the gap between how well something works in training and how it works in the world
What the class is really about
This is the class where the idea becomes physical, because you build the thing yourself. You gather your own examples, train a model on them, and test it. Getting something working takes about ninety seconds. The rest of the class is spent finding out where it breaks.
The breaking is the lesson. A model trained on thirty photographs of your own hand will recognise your hand and then fail on the hand of the person sitting next to you. Nobody has to explain bias to a learner after that has happened in front of them. They watched where it came from. It came from what they showed it.
Three steps carry the whole idea. Gather examples. Train. Test. Then do it again with better examples. What a learner takes away is that a model is a mirror of the material it was handed. It did not learn the idea of a hand. It learned the particular patterns in the particular pictures it was given, and it cannot tell the difference between those two things.
From there the real examples land properly. Face recognition that works better on some faces than others. Hiring tools that carry forward the habits of the hiring they were trained on. Voice assistants that struggle with an accent they were not shown enough of. None of that is a machine deciding to be unfair. The material was not neutral, so the result is not neutral either.
You also see why more data of the same kind is worth less than a smaller amount of varied data, which is the opposite of what most people assume.
Careful use
Bias. Where it comes from, and why it is not the machine being unfair on purpose.
Careful use is not a separate lesson at the end of the course. A piece of it belongs to every one of the eight classes, and this is the piece that belongs to this one.
How this class is taught
- One core concept. The class carries a single idea. Nothing is covered twice and nothing is rushed.
- Hands before theory. You do the thing first. The explanation lands better once you have seen what happened.
- Practice in company. Weekday slots in the lab follow the weekend class, with somebody there to help.
- Mixed ages at one table. The youngest in a batch may be ten and the oldest past fifty. Each takes something from the other.
Chapter 2, Prompt Writing · Chapter 4, Breadboard Fundamentals