What Is Machine Learning? A Visual Introduction for Beginners

Your email app sorts millions of messages into spam and not spam. Yet nobody sat down and wrote a rule for every scam ever invented. So how does it know?

• Spam or not spam?
• Nobody wrote every rule

Imagine trying to write those rules by hand. Scammers change their wording daily, and every new trick breaks your old rule. The better idea is to let the computer find the patterns itself, from examples.

• Hand-written rules break
• Learn patterns from examples

Welcome to episode one of Machine Learning from Zero to Advanced. Today we build the big picture: what machine learning is, the words you need, the main types, and how a real project flows.

• Big picture first
• No heavy math today

Let's untangle some buzzwords. Artificial intelligence is the broad goal of making machines do things that seem intelligent. Machine learning is a part of that, where systems learn from data instead of following fixed rules. Deep learning is a part of machine learning that uses large neural networks.

• AI: the broad goal
• ML: learning from data
• Deep learning: big neural networks

Data science is a neighbour that overlaps all of this. It uses statistics, programming and domain knowledge to answer questions with data, and machine learning is one of its most powerful tools.

• Data science overlaps ML
• Statistics plus code plus context

Here is a simpler example to hold in mind. You want to predict the price of a house. You have records of past sales, each with details like size and number of bedrooms, and the price it sold for.

• Past sales as examples
• Predict a new price

The details we feed in, like size and bedrooms, are called features. The thing we want to predict, the price, is called the label. Each row is one example the model can learn from.

• Features: the inputs
• Label: the answer
• Row: one example

A model is a flexible recipe with adjustable dials. Those dials are the parameters. Given features and the current dial settings, the model produces a prediction, its best guess at the label.

• Parameters: adjustable dials
• Prediction: the model's guess

How do we know whether a guess is good? We compare the prediction to the true label and compute a loss, a single number that says how wrong the model is. Small loss means good, large loss means bad. Everything in learning is about pushing that number down.

• Loss: one number for error
• Learning means lowering it

Now the main types of learning. In supervised learning, every training example comes with the correct answer. Spam filtering and house price prediction are both examples: the model learns from emails already marked spam, or from houses with known prices.

• Examples come with answers
• Example: spam filter

In unsupervised learning there are no answers. The model just looks for structure on its own. A shop might feed in customer purchase data and let the model discover groups of similar shoppers.

• No answers provided
• Example: customer groups

Self-supervised learning creates its own answers from the data. Hide a word in a sentence and ask the model to guess it. The hidden word is the label, and no human had to write it. This idea powers modern language models.

• Hide part, predict it
• Example: guess the missing word

Reinforcement learning is learning by trial and reward. An agent takes actions, sees what happens, and gets rewards or penalties. Think of a program learning to play a game, getting better with every attempt.

• Agent, action, reward
• Example: learning a game

Now let's watch a model actually learn. Take the house data and plot it: size on the horizontal axis, price on the vertical. The points drift upward, bigger houses tend to cost more.

• Size versus price
• A clear upward trend

Our model is a straight line, and its parameters control the tilt and the height. We start with a random setting, and the line is a poor fit. Vertical bars show each miss, and together they give a high loss.

• Line = the model
• Red bars = errors
• Loss starts high

Learning is a loop. Nudge the parameters a little in the direction that reduces the loss, check again, and repeat. The line tilts and slides, the red bars shrink, and the loss falls step by step until the line sits snugly through the cloud.

• Nudge parameters
• Loss falls each step
• Line fits the data

Now the model can predict. Give it the size of a house it has never seen, and it reads the price off the line. That is the heart of machine learning: adjust parameters to fit examples, then use them on new cases.

• Fit on known examples
• Predict on unseen cases

Real projects follow a workflow. First, define the problem clearly: what are you predicting, and how will you judge success? Second, gather and clean the data, because a model can only learn what the data shows.

• Define the problem
• Collect and clean data

Next, split the data. Most goes to training, and a portion is hidden away for testing. Train the model on the first part, then evaluate it on the hidden part to see whether it works on data it has never seen.

• Train on most of the data
• Test on hidden data

If the results are good, deploy the model so that it serves real users. The story doesn't end there, because the world changes, so you monitor the model and retrain it when it drifts.

• Deploy to real users
• Monitor and retrain

Here is the journey ahead. We start with the mathematics: algebra, linear algebra, calculus, and probability, all from scratch. Then Python and data skills, then models like regression, trees and support vector machines, and finally neural networks, transformers and reinforcement learning.

• Math from scratch
• Classic models
• Deep learning and beyond

Our tools are simple. Python is the language, Jupyter notebooks let us run code and see results side by side, and NumPy gives us fast arrays and the vector maths machine learning runs on. You will also implement algorithms from scratch, so you understand them, not just use them.

• Python
• Jupyter notebooks
• NumPy

To recap: machine learning lets computers learn patterns from data instead of following hand-written rules. Features go in, the model's parameters produce a prediction, and loss measures the error. We met supervised, unsupervised, self-supervised and reinforcement learning, and the workflow from problem to deployment.

• Learn patterns from data
• Loss guides learning
• Four types, one workflow

This was episode one. Next time we begin the mathematics with algebra essentials: functions, exponents, logarithms and sigma notation, the language every model speaks.

• Next: algebra essentials
• Subscribe for the series