What Is Machine Learning? Definitions from Samuel, Mitchell & Experts

The simplest answer is:

Machine Learning is a way of building computer systems that learn useful patterns from data or experience and use them to make predictions or decisions without us explicitly programming every rule.

But there isn’t one universally accepted definition.

Over the history of Machine Learning, researchers have emphasized different ideas: learning from experience, improving at a task, finding patterns in data, and generalizing to situations the system hasn’t seen before.

Two definitions in particular have shaped how Machine Learning is taught today: those associated with Arthur Samuel and Tom Mitchell.

Let’s jump right in!

Arthur Samuel’s Definition of Machine Learning

Arthur Samuel was an IBM researcher and an early pioneer of Artificial Intelligence. He is credited with popularizing the term Machine Learning in 1959 through his work on a computer program that learned to play checkers.

The definition commonly associated with Samuel describes Machine Learning as:

A field of study that gives computers the ability to learn without being explicitly programmed.

There’s an interesting historical detail here.

Although this sentence is almost universally attributed to Samuel, the exact wording does not appear in his famous 1959 paper Some Studies in Machine Learning Using the Game of Checkers.

The underlying idea, however, absolutely does.

Samuel was investigating whether a computer could improve through experience rather than requiring a programmer to explicitly specify every possible behavior.

His checkers program is a perfect example.

Instead of programming rules for every possible board position, Samuel created a system capable of using experience to improve how it played.

This gives us the first important idea behind Machine Learning:

We don’t have to explicitly program every decision rule ourselves.

Traditional Programming vs Machine Learning

This distinction becomes clearer if we compare Machine Learning with traditional software.

In traditional programming, we usually have:

Data + Rules → Answers

A programmer writes the rules.

The computer executes them.

Imagine a very simplistic spam filter:

IF email contains "FREE MONEY" → SPAM
IF email contains 10 links → SPAM
IF sender is blacklisted → SPAM

We could keep adding rules.

But reality quickly becomes too complicated.

Machine Learning approaches the problem differently.

For supervised learning, we can provide:

Data + Known Answers → Learned Model

For example, thousands of emails labelled:

SPAM

or:

NOT SPAM

The algorithm uses those examples to learn patterns associated with each category.

We can then apply the resulting model to emails it has never seen before.

This is much closer to what Samuel meant by learning rather than explicitly programming every possible situation.

Tom Mitchell’s Definition of Machine Learning

Almost four decades later, computer scientist Tom Mitchell provided a much more formal definition.

In his influential 1997 textbook Machine Learning, Mitchell describes learning in terms of three components:

Experience (E)
Task (T)
Performance measure (P)

In Mitchell’s formulation, a computer program learns when its performance at a task T, measured by P, improves through experience E.

This definition might look academic at first.

It’s actually extremely practical.

Understanding E, T and P

Let’s return to our spam classifier.

T — Task

What do we want the system to do?

Classify new emails as spam or not spam.

That’s our task.

E — Experience

What does the system learn from?

For example:

100,000 historical emails labelled spam/not spam.

That’s its experience.

P — Performance

How do we know whether it has become better at the task?

We need a metric.

Depending on the problem, that could be:

Accuracy
Precision
Recall
F1 Score
ROC AUC
Another appropriate evaluation metric

If the model’s performance on new emails improves as a result of learning from those historical examples, then according to Mitchell’s definition:

the program has learned.

Another Example: Learning to Play Chess

Mitchell’s framework can be applied to almost any ML problem.

Imagine a system learning to play chess.

T — Task: Play chess.

E — Experience: Games played against humans, other programs or itself.

P — Performance: Percentage of games won, Elo rating, or another measure of playing strength.

If additional experience causes its measured chess performance to improve, the system is learning.

This is why Mitchell’s definition has remained so useful.

It forces us to answer three fundamental questions:

What are we trying to do?

What are we learning from?

How will we know whether we’re getting better?

Those questions are just as relevant to modern Machine Learning as they were in 1997.

A Modern Definition of Machine Learning

Today, I find it useful to think about Machine Learning like this:

Machine Learning uses algorithms to learn patterns from data or experience that generalize to new observations.

The word generalize is particularly important.

A model isn’t useful simply because it remembers its training data.

Imagine a student who memorizes every answer to a practice exam.

They score:

100%.

Great.

Now change the questions slightly.

They fail.

Did they really learn?

Machine Learning models can have exactly the same problem.

It’s called overfitting.

A model can perform extremely well on the data it learned from while performing badly on new observations.

The real objective is therefore not memorization.

It’s generalization.

RECOMMENDED PATH

The Fastest Way to Lean ML

Follow a structured 90-day plan to go from zero to real-world projects.

Start Free 90-Day ML Roadmap →

What Do Machine Learning Definitions Have in Common?

Despite differences in wording, most useful definitions of Machine Learning contain some combination of five ideas.

1. Experience or Data

The system needs something from which to learn.

That could be:

labelled examples
unlabelled observations
interactions with an environment
human feedback
historical data

2. A Task

We normally want the system to accomplish something.

For example:

classify an image
predict a house price
forecast demand
detect fraud
recommend a product
generate text

3. Learning

Some aspect of the system changes based on its experience.

4. Performance

We need a way to determine whether the system is actually useful at its task.

5. Generalization

The learned patterns should work beyond the exact examples used during training.

Put everything together and we get:

Experience → Learning → Model → New Data → Prediction or Decision

That’s a useful mental model for Machine Learning.

Machine Learning vs Artificial Intelligence

Machine Learning and Artificial Intelligence are related, but they’re not synonymous.

Artificial Intelligence (AI) is the broader field concerned with building systems capable of performing tasks associated with intelligent behavior.

Machine Learning (ML) is one approach to building those systems.

A simplified hierarchy is:

Artificial Intelligence

↓

Machine Learning

↓

Deep Learning

Deep Learning is therefore a subset of Machine Learning, which itself sits within the broader field of AI.

Deep Learning vs Machine Learning

It’s also worth remembering that Machine Learning doesn’t necessarily mean ChatGPT-style Generative AI.

A simple logistic regression model predicting whether a customer will churn is Machine Learning too.

The Main Types of Machine Learning

The definition becomes even clearer when we look at how machines can learn.

Supervised Learning

The algorithm learns from examples where the desired answer is known.

Examples:

Spam vs not spam
Fraud vs legitimate transaction
House-price prediction
Disease classification

Two of the most common supervised learning tasks are classification and regression.

Unsupervised Learning

The algorithm receives data without a known target label and attempts to identify useful structure.

For example:

Customer segmentation
Clustering similar documents
Finding unusual observations
Discovering patterns in high-dimensional data

Reinforcement Learning

An agent learns by interacting with an environment and receiving feedback through rewards.

Games are an intuitive example.

The agent tries different actions and gradually learns which behaviors tend to produce better outcomes.

So, What Is the Best Definition of Machine Learning?

There isn’t one perfect definition.

The definition associated with Arthur Samuel captures one of the field’s revolutionary ideas:

computers can learn behavior without us explicitly programming every rule.

Tom Mitchell gives us a more rigorous framework:

Experience + Task + Performance.

And a modern practical definition emphasizes another crucial idea:

generalization to new data.

Put those perspectives together and I’d define Machine Learning as:

The study and application of algorithms that improve at a task through data or experience and generalize what they learn to new situations.

That’s the idea behind everything else you’ll encounter in Machine Learning.

Regression, decision trees, neural networks, training, validation, features and evaluation are different pieces of machinery built around it.

Now That You Know What Machine Learning Is, What Should You Learn Next?

This is where learning ML often gets confusing.

Should you start with Python?

Statistics?

Linear regression?

Scikit-learn?

Neural networks?

You don’t need to learn everything at once.

Our Free 90-Day Machine Learning Roadmap organizes the journey from Python and mathematics through the core ML algorithms, model evaluation and your first real projects.

→ Start the Free 90-Day Machine Learning Roadmap

RECOMMENDED PATH