# Your alphabet

Part 2 of the tutorial "neural-networks".
Canonical: https://learn.welldun.ai/neural-networks/02-your-alphabet/

Invent characters that have never existed and draw a dataset of them by hand. Nothing here mentions yet what the network will receive.

---

Make up a few characters and draw each one a handful of times. A hook. A spiral
with a tail. A box with a dot inside. They only need to look different enough
that you could tell them apart yourself.

Aim for three characters with about eight drawings of each, and draw them a
little differently every time — bigger, smaller, leaning, rushed. You can come
back and add more whenever you like.

Two characters will work, but everything downstream is more interesting with
three: with only two, the network's scorecards are forced to be exact mirror
images of each other, so several of the later parts end up showing you the same
picture twice.

> **[interactive figure: alphabet-builder]** — needs a browser; see https://learn.welldun.ai/neural-networks/02-your-alphabet/

## Why draw a dataset by hand?

What you are making is a **dataset**: a pile of examples, each one labelled
with the right answer. The label is the whole point. Without it there is
nothing to be wrong about, and being wrong is the only way this network learns
anything.

Because you invented the characters, nothing out there has ever seen them. No
pretrained model can help, and nothing can be looked up. Every number the
network ends up with will have come from your hand.

Drawing each character several times, differently each time, is not busywork
either. Part 06 shows exactly what the variation buys you — the parts of your
handwriting that change from drawing to drawing cancel each other out, and what
survives is the character itself.

Your drawings stay in this browser. There is no account and no server.

If you want to see what real handwriting datasets look like:
[Omniglot](https://github.com/brendenlake/omniglot) collects 1,623 characters
from 50 alphabets, [MNIST](https://en.wikipedia.org/wiki/MNIST_database) is
70,000 handwritten digits and the standard first exercise, and
[EMNIST](https://www.nist.gov/itl/products-and-services/emnist-dataset)
extends it to letters. Yours will be about twenty drawings — which turns out
to be enough, because you invented characters that look nothing alike.