Watch it fail

Here is the machine, the four wine types, and the same delta rule from the last chapter. All four weights start equal at 20, exactly as they started equal at 8 before.

Press correct it. Each press corrects the weights on the bottle in the machine, then puts the next of the four wines in front of it.

Go round all four and run 25 appears, which repeats that loop twenty five times over. One set of weights meets each of the four wines in turn, and is wrong about every one of them.

Watch two things.

The total error. It starts at 60 dollars. Press once and it is still 60. Run twenty five more passes and it is still 60. It does not creep down. It does not bounce. It sits there.

Compare that to the last chapter, where one pass took 114 dollars down to 69, and the next took it to 49. Something was clearly happening. Here, nothing is.

The four weights. These do move, and where they move to is the interesting part. They drift toward each other, and they end up all but identical, at around 12 and a half each.

Think about what identical weights mean. Back at the start of the last chapter, four equal weights was how the model said I have no idea which of these matters more. It began in that state because it had seen nothing.

This model has now seen hundreds of bottles, and it has arrived back at exactly the same statement.

Then look at the four guesses underneath. Every wine, whatever grape, whatever flavour, gets priced at almost exactly 25 dollars.

Twenty five is not a random number. The wines cost 40 or 10, and 25 sits precisely between them. The model has worked out that it cannot tell these bottles apart, so it is quoting the average and taking the loss on every single one.

That is not a model that is learning slowly. That is a model that has finished, and this is its best answer.