Okay... So, What Exactly Is Machine Learning Again?

If you've spent even a few minutes reading about technology lately, you've probably seen the term AI everywhere.

At least for me, it feels like every LinkedIn profile I come across has the title 'AI Engineer.' AI has become the buzzword of the decade.

But if there's one concept worth understanding before diving into the world of AI, it's Machine Learning.

Why?

Because Machine Learning is the engine behind many of the 'smart' experiences we use every single day.

Think about it.

How does Netflix seem to know what you'll binge next? Why does Gmail catch spam before you even see it? How does Google Maps predict traffic? Or how does your phone unlock simply by recognizing your face?

These features have become such an integral part of our digital experience that we rarely stop to think about them. They're so seamless that it's easy to forget that systems capable of making intelligent predictions existed long before the rise of Large Language Models and generative AI.

Which brings us to an important question:

Where exactly does Machine Learning fit into the AI landscape? Is it the same as AI? Is it a subset?

To answer those questions, we first need to understand what Machine Learning actually is.

What Is Machine Learning?

If you search for 'What is Machine Learning?' on Google, you'll probably come across something like this:

Wikipedia gives a slightly more detailed explanation:

These definitions are technically correct, but if you're new to the field, they can feel a little abstract. Terms like algorithms, statistical models, and predictions don't necessarily make the concept click.

So instead of starting with formal definitions, let's build some intuition.

Because once the intuition makes sense, the definitions become much easier to understand.

A Simple Analogy

Suppose you're a programmer, and the only program you've ever written is this:

y = x + 10

It's simple.

Give the program an input x, and it always returns x + 10.

If x = 5, the answer is 15.

If x = 20, the answer is 30.

In this case, you already know the rule. Your job as the programmer is to write that rule in code, and the computer simply follows your instructions.

This is how traditional programming works.

For decades, we've solved problems by designing algorithms, a sequence of well-defined steps and translating those steps into code. Computers don't ‘think’; they execute the instructions we give them, only much faster and more accurately than we can.

And honestly, that's pretty amazing. This simple idea is what enabled everything from calculators and operating systems to banking software and space missions.

But now let's flip the problem.

Imagine I hand you a list of numbers:

This time, I don't tell you the rule.

All I give you is the data.

Looking at the table, you can probably guess that the relationship is:

Output = Input + 10

Easy enough.

But that's only because the pattern is obvious.

Now imagine the table contains millions of rows instead of five.

Instead of numbers, it contains images, customer purchases, emails, medical reports, sensor readings, or even human conversations.

The relationship between the inputs and outputs is no longer obvious. In fact, it may be impossible for a human to write down all the rules by hand.

So what if, instead of us discovering the rule...

...we could build a system that discovers the rule for us?

That's the central idea behind Machine Learning.

Rather than programming the rules explicitly, we provide the computer with examples and let it learn the underlying patterns from the data.

In other words:

This is exactly what Machine Learning is all about.

But here's the twist.

What if I told you that Machine Learning isn't what it’s sold for?

At its core, it's still made up of algorithms—the same kind of algorithms we've been writing for decades.

Sounds a little underwhelming, doesn't it?

The difference isn't that we invented an entirely new way of programming. We simply changed what the algorithm is responsible for.

In traditional programming, we design the rules ourselves. We carefully think through the problem, write the logic, and ask the computer to execute it.

In Machine Learning, we flip that responsibility.

Instead of writing the rules, we choose a learning algorithm and provide it with data. The algorithm then analyzes the data, searches for patterns, and builds a model that captures those patterns.

When we say an algorithm 'learns,' we mean that it adjusts its internal parameters to find patterns that help it make better predictions.

In other words, we no longer tell the computer what the rules are, we ask it to discover the rules for us.

Of course, this doesn't mean the computer magically ‘understands’ the data. Under the hood, these learning algorithms are performing a tremendous amount of mathematics and statistics. They compare examples, measure errors, optimize parameters, and gradually improve until they find a pattern that best explains the data.

That's why you'll often hear another term used interchangeably with Machine Learning: Statistical Learning.

Machine Learning is, in many ways, applied statistics at scale. Instead of manually analyzing data to uncover relationships, we build algorithms that can do that automatically, even when the data contains millions of examples and hundreds or thousands of variables.

So what exactly does the algorithm produce after learning from all that data?

It produces a machine learning model.

Machine learning models can be thought of as templates that learn a set of rules from data. We provide the model with examples, and it learns patterns that allow it to map inputs to outputs

The 'model' is not a separate entity from the algorithm, it is the algorithm after it has learned from data. During training, the algorithm adjusts its internal parameters so that it better fits the data it has been given.

The better the model fits the underlying patterns in the data, the more accurate its predictions are likely to be. However, we do not aim for a perfect (100%) fit to the training data. A model that memorizes every training example may perform poorly on new, unseen data, a problem known as overfitting. Instead, the goal is to learn the general patterns in the data so the model performs well on data it has never seen before

Training a machine learning model involves feeding it a large amount of data and allowing it to repeatedly adjust its parameters to reduce prediction errors. Although the idea is conceptually simple, the underlying mathematical computations are computationally intensive. This is why training modern machine learning models often requires powerful hardware.

Once training is complete, the model has learned a function that can map inputs to outputs. At this stage, we can provide it with new inputs (in the same format as the training data), and it will generate predictions. This stage is called inference or prediction.

Since machine learning relies heavily on probability and statistics, the model's predictions are usually the most likely outputs based on what it learned during training rather than guaranteed correct answers. As a result, every model has some degree of error.

A model achieving 90–95% accuracy can be considered excellent for many applications, but whether that is 'good enough' depends entirely on the problem. 

Example: A spam email detector with 95% accuracy may be perfectly acceptable, while a medical diagnosis system or a self-driving car would often require much higher reliability.

So, in short machine learning models make predictions based on patterns they have learned from data. They are not perfect, but when trained well and evaluated appropriately, they can make highly reliable predictions for many real-world tasks.

Where Does It All Fit?

Now that we have an intuition for what Machine Learning is, let's answer the question we left hanging earlier:

Where does Machine Learning fit in the world of AI?

Artificial Intelligence is the broader field, the ambitious goal of building machines capable of performing tasks that typically require human intelligence

Machine Learning is one of the most successful approaches to achieving that goal. Instead of programming every rule by hand, we build systems that learn those rules from data.

And then there's Deep Learning, a specialized branch of Machine Learning that uses neural networks with many layers to learn increasingly complex patterns. But all that for later

For not a simple hierarchy looks something like this:

Understanding this hierarchy is important because throughout this series we'll mostly focus on Machine Learning—the foundation upon which many modern AI systems are built. Once you're comfortable with these fundamentals, concepts like Deep Learning and Large Language Models become much easier to understand.

And that's Machine Learning in a nutshell.

Behind all the buzzwords, it's simply a different way of solving problems. Instead of telling a computer exactly what to do, we give it examples and let it discover the patterns for itself.

Once you understand this idea, the rest of the Machine Learning landscape becomes much less intimidating. In the next article, we'll explore the different ways machines learn and why choosing the right approach depends on the problem you're trying to solve.