Good ingredients make good AI: data

Garbage in, garbage out

A machine learns only from the examples we give.
So if the examples are wrong,
it faithfully learns those mistakes too.
Label a cat photo as a dog,
and the machine will insist a cat is a dog.
Give it only one-sided examples,
and it believes the world leans that way.
From material that even a person sees as junk,
even the best machine learns only junk.

Feed in the wrong examples one by one. The more mislabeled data goes in, the more the machine's judgment goes astray.

As wrong examples piled up,
the machine's answers fell apart.
The machine cannot tell right from wrong on its own
and simply trusts the material we give.
So mislabeled or skewed material
turns straight into a bad habit.
Put another way, just fixing the data
makes the same machine far smarter.

Good ingredients, proper learning

So what makes a good ingredient?
First, the labels must be accurate.
A cat as a cat, a dog as a dog.
Second, it must not lean to one side.
Give heaps of cat photos and only a few dogs,
and the machine barely recognizes dogs.
Show both sides evenly and it learns in balance.
Third, there should be little noise muddling the same thing.
Blurry or oddly mixed material is best left out.

Tap to swap the junk data for clean, balanced data. The moment the ingredients improve, the machine's accuracy rises.

Without touching the machine at all,
the accuracy shot up.
The only thing that changed was the ingredients.
Fix the labels and fill both sides evenly,
and the same machine learned properly.
That is why, before tinkering with the model,
people pour effort into cleaning the data first.
Good ingredients are half of a good result.

Pick features that fit the problem

Even with clean data, one thing remains.
Deciding what it should look at to judge.
The clues we pick out are called features.
Think of the problem of guessing house prices.
A house's area or location is deeply tied to its price.
Big, in a nice neighborhood, the price goes up.
But what about the color of the door?
It has almost nothing to do with the price.
Such an irrelevant feature only clouds the judgment.

Pick the features for predicting house prices. Add area and location and accuracy rises; add an irrelevant one like door color and it drops instead.

Each fitting feature you picked raised the accuracy.
The irrelevant color, by contrast, dropped it.
More features is not automatically better.
You must pick clues truly tied to the problem.
A fitting feature is a shortcut to the answer,
while a wrong one is noise that confuses the path.
Choosing well which clues to show,
this too is important work in making good data.

Good data is better in plenty

If the ingredients are clean and the features fit,
now the amount starts to matter.
With only a handful of examples,
the machine is easily swayed
by one or two odd cases that happened to slip in.
With very many examples,
such flukes get buried in the average and weaken,
while the real, frequent pattern grows clear.
So clean data is more reassuring in plenty.

Tap to add more data. As the dots grow many, chance wobble settles, the boundary steadies, and accuracy rises.

The more data you added,
the steadier the boundary and the higher the accuracy.
When scarce, even one point made it lurch,
but as it grew, the wobble died down.
Of course, just adding amount is not the answer.
No matter how much garbage you gather, it is still garbage.
Only when the data is clean and unbiased
does more truly mean better.
Quality comes first, then quantity.

Let's wrap up

Gathered on one line, it is this.
A machine learns only as well as the data we give.
Garbage in, garbage out,
but clean, balanced ingredients let it learn properly.
You must pick features that fit the problem,
while irrelevant clues only get in the way.
Good data is more reassuring in plenty,
but quality first, quantity second.
Good ingredients make good AI.

Tap the key points in order to review. (garbage in, garbage out -> clean ingredients learn right -> pick fitting features -> good data, the more the better)

Now you know why data matters so much.
People often picture a cleverer model,
but what truly decides the result is often the ingredients.
Clean it up, gather it without bias,
pick features that fit the problem, and give it in plenty.
This plain-looking work is the foundation of good AI.
Next, with data prepared this way,
let's keep exploring how we gauge
how well the machine actually learned.