Eight bottles, one at a time

Here is the training set, and here is the loop. Nothing new to learn, just the same four steps running over and over.

Press correct it and one bottle goes through. Its weights get nudged, and the next wine arrives already carrying them. Eight presses takes you through the whole set.

The ember numbers above the weights are the coming correction, shown before you commit to it. That is error times unit times learning rate, the same three numbers as ever, worked out in advance so you can see the nudge before it lands.

Bottle six is worth slowing down for. Every wine so far has been guessed too cheaply. That one is the first the model guesses too expensively, and the next slide is about what the rule does when that happens.

When the eighth bottle is done, read the summary that appears. One row per wine, and three prices across each row: what the model guessed at first, before any correction; what it guesses now; and the actual price it was always aiming at.

Compare the middle column against the right one. Seven of the eight have moved closer. The one that has not is marked with a small arrow. Underneath, the same thing said once: add up how far every guess sits from its real price, ignoring whether it was too high or too low, and that total has come down from about 114 dollars to around 69.

Not one bottle is priced correctly, and one is worse than when it started. The set as a whole is better. Both are true, and the second is the one that matters.

Go round again starts a fresh round. The same eight wines arrive in the same order, beginning with bottle one, but the weights are not reset. The model meets that first bottle carrying everything the previous round taught it.