Learn machine learning · before you start
Before you start
Five drawings, in the order the track uses them. If you have never met machine learning, this is the whole vocabulary you need to open session 1.1 and follow it.
A table of examples
One row is one application
illustrativeOur first example is a . One row is one loan application, each column is one thing you know about it at the moment you have to answer, and the last column is what happened afterwards: repaid, or not repaid. The applications are made up by a script running in your browser, so nobody real is in them and no money is at stake.
A prediction
Cover the answer, write a guess, lift the cover
illustrativeThe last cell is covered, so the answer cannot be read.
A is that cover and that guess. Do it for two thousand rows, count the ticks, and you have measured something: the fraction you got right is called accuracy. Later you will use other scores that measure different kinds of mistakes, including how confident a wrong prediction was.
A model
Columns in, one guess out, a dial on the side
illustrativeA is whatever sits in that box: columns go in, one guess comes out, and there is at least one dial on the side that changes what comes out. For now the box can stay shut. You will build what is inside it in Parts 2 to 10, starting with a box that is three questions long.
Training
Turning the dial until more rows come out right
illustrativeRows with their answers go in. Two of the six drawn rows come out right.
is what turns the dial. I feed the box rows whose answers I already have, compare what it says against what happened, and adjust to improve a chosen score on the training examples. To check how well that carries over, you need examples that were not used to fit the model.
Held-out rows
One pile to work on, one pile sealed
illustrativeTwo thousand synthetic applications in a single pile.
So the applications are cut in two before anything runs. Fourteen hundred are yours to work on, and the other six hundred go into a sealed pile: . They help estimate performance on new applications drawn in the same way. The training score alone can be too optimistic. When you start choosing settings, you will use a validation pile for those choices and keep a separate test pile for the final check. Later lessons use different datasets and split sizes.
How a session runs
Read the question, then predict what will happen. Try the controls and compare the result with your prediction. Check your understanding by reaching a target, answering questions on new examples or demonstrating the idea. Finally, save your takeaway in the notebook.
A wrong prediction costs nothing. Your answers stay in this browser. Writing down a guess gives you something to compare with the result, so you can see which part of your explanation needs to change.