Four pillars of machine learning

The four pillars of machine learning are data, a model, training, and testing. Data is the examples the computer learns from. The model is the math that finds patterns in those examples. Training is the practice rounds where the model guesses, gets corrected, and adjusts. Testing is checking the model on stuff it has never seen, so you find out if it actually learned or just memorized.

That's it. Everything else is detail.

Different people slice this up different ways. Some folks call the pillars data, algorithms, compute, and evaluation. Same idea. There's no official list carved in stone somewhere. What I care about is that you understand what each piece does, because every AI tool you've heard of (ChatGPT, Gemini, the AI answer at the top of Google) sits on top of these four.

What is machine learning, in plain words?

Machine learning is a way of teaching a computer by showing it examples instead of writing out every rule by hand.

Say you wanted a computer to spot junk email. The old way, somebody writes a rule: if the email says "free money," block it. Then scammers write "fr3e m0ney" and the rule breaks. The machine learning way, you show the computer a pile of emails that are junk and a pile that aren't, and it works out the pattern on its own.

It's a tool, like a calculator. A calculator doesn't understand math. It follows steps really fast. Machine learning doesn't understand your business either. It finds patterns really fast. Keep that in your head and a lot of the hype goes away.

Pillar one: data

Data is the examples. Text, photos, prices, reviews, web pages, whatever the thing is supposed to learn from.

This is the pillar that matters most, and it's the one people skip. A model can only learn what's in the data. If the data is thin, it learns thin. If the data is wrong, it learns wrong, and it'll be confident about it.

Think about training a new hire by handing them a binder. If half the binder is outdated and the other half contradicts itself, you don't have a bad employee. You have a bad binder.

That's the part that should matter to you as a business owner. The information about your business floating around online is data. Your website, your Google Business Profile, your reviews, your listings. If your hours say one thing on Google and another on Facebook, that's a contradiction in the binder. I wrote more about that in why consistent business information matters.

Pillar two: the model

The model is the pattern-finder. It's a set of math that takes something in and gives an answer back.

You'll hear the word "algorithm" here too. An algorithm is just a recipe, a list of steps. The model is what you get after you run that recipe on a bunch of data. Before training, it's a blank recipe card. After training, it's a recipe card with the measurements filled in.

There are a lot of kinds of models. Some are simple and sort things into two buckets, like junk or not junk. Some are huge, like the ones behind ChatGPT, which are called large language models. That's a fancy name for a model trained on a massive amount of text so it can predict what words come next.

You don't need to know how the math works. You need to know the model only knows what it was fed, which loops right back to pillar one.

Pillar three: training

Training is practice. The model looks at an example, makes a guess, and gets told how wrong it was. Then it adjusts a tiny bit. Then it does that again, over and over, a huge number of times.

It's like learning to shoot free throws. You miss left, you adjust. You miss short, you adjust. Nobody hands you the perfect shot. You get there by missing a lot.

Training is also where the cost lives. It takes a lot of computer power, which people call "compute." That's why the big AI models come from big companies. Running that many practice rounds on that much data isn't cheap.

One catch worth saying out loud: training happens at a point in time. The model gets trained, and then the world keeps moving. That's why a lot of AI tools now also go look things up on the web when you ask a question, instead of relying only on what they learned in training. So your business can show up two ways. It can be in what the model learned, or it can be in what the tool finds when it goes looking.

Pillar four: testing

Testing is checking the model on examples it has never seen. People also call this evaluation.

This pillar exists because a model can cheat without meaning to. If you quiz it on the same examples it practiced on, it can look perfect just by memorizing the answers. That's the student who memorized the practice test and failed the real one. So you hold some examples back, and you only use them at the end to see if the model really learned the pattern.

If it does well on the new stuff, you've got something useful. If it doesn't, you go back to the other three pillars. A lot of the time the fix is better data, not a fancier model.

How do the four pillars work together?

They're a loop, not a checklist.

• Data goes in.
• The model is the thing being shaped.
• Training shapes it by guessing and correcting.
• Testing tells you if it worked, and sends you back to the start if it didn't.

Pull any one of them out and the whole thing falls over. Great data with no testing, and you don't know if it works. A great model with bad data, and it learns the wrong lesson really well.

Why should a business owner care about any of this?

Because customers are asking AI tools who to hire now, and those tools are built on these four pillars.

You can't touch the model. You can't touch the training. You can't run the testing. The only pillar you have any say in is the data, meaning what's out there about your business and how clear it is. Your job is to make that information easy to read, easy to trust, and the same everywhere.

A lot of that comes down to how your website is written and built. Plain answers to real questions. Clear service pages. Your name, address, and phone number matching everywhere they show up. If you want to see how an AI tool actually goes through a page, read how AI reads website content.

I'm not going to tell you that fixing your data guarantees you show up. Nobody can promise that, and anybody who does is guessing. What I can tell you is that a model can't recommend a business it barely has any information on.

Where to start

Start by checking what AI tools say about you right now. Open ChatGPT in incognito mode, so there's no history steering the answer, and search for your service in your city. See who comes up. I walk through the whole thing in how to test if AI can find your business.

If you'd rather have somebody handle the cleanup, that's the work we do at Heaston Innovations. You can see what that looks like on our services page.

At the end of the day, machine learning is four pieces: data, model, training, testing. Three of them belong to the big AI companies. One of them is yours.