Read my writing

The network below has been trained on your drawings for 400 rounds, using the rule from the last part. Nothing else about it is special. Draw one of your characters and it will tell you which one it thinks you drew.

What happens if you draw something it was never taught?

Draw a spiral, a house, a letter from a real alphabet — anything at all. It will still answer with one of your characters, often with great confidence. It has no category for “none of these” and no way to invent one.

That is not a fault in your network; it is what a classifier is. It was asked to divide the world into a fixed number of boxes, so it divides the world into that many boxes. Every one of these systems has an edge like this, and knowing where the edge sits is most of using them responsibly.

How much should you trust the accuracy figure?

Some of your drawings were kept out of training entirely. The score on those is the only one worth anything, because they are the ones it could not have memorised — which is why the figure above reports them separately.

It has seen a few dozen drawings in its life, all by one person, all with the same pen. If your handwriting drifts, or someone else draws your characters, expect it to struggle. More drawings, and more varied ones, is the cure for almost everything that goes wrong here.

What have you actually built?

You invented an alphabet nobody had written before, turned it into numbers, built a network out of nothing but a table and a multiply-and-add, worked out how wrong it was, and taught it to be less wrong — and now it reads your handwriting.

There was never anything in the box but numbers being nudged.

The same rule, stacked into layers and pointed at photographs, is image recognition. Pointed at sequences of characters, so it predicts the next one rather than naming the whole, it becomes a language model. Nothing new gets added — only more layers, more numbers, and a great deal more data.

If you want to keep going: draw more characters and watch it get harder, add a third and a fourth, or go back and see what happens when you train on four drawings instead of all of them.