Skip to content
AI360Xpert
Core ML
Visual explainer

What Is Machine Learning?

From hand-written rules to learning patterns from examples. The paradigm shift that makes AI possible.

Writing rules by hand fails when the task gets too complex
Writing rules by hand fails when the task gets too complex

For decades, programming meant writing explicit instructions. If you wanted to detect a cat in a photo, you had to write a rule for every pixel, ear shape, and shadow. It failed because the real world is too messy for if/else statements.

The paradigm shift

Instead of writing rules, we provide examples and let the machine learn the mapping from X to Y
Instead of writing rules, we provide examples and let the machine learn the mapping from X to Y

Machine learning flips the approach. Instead of giving the computer the rules, you give it the answers (labels, YY) and the data (inputs, XX). The algorithm finds the rules that map XX to YY automatically.

Three ways to learn

Supervised uses labels, unsupervised finds structure, and reinforcement learns from rewards
Supervised uses labels, unsupervised finds structure, and reinforcement learns from rewards

Supervised learning gets explicit answers. Unsupervised learning just gets raw data and has to find the hidden structure itself. Reinforcement learning learns by interacting with an environment and collecting rewards or penalties over time.

The true test

Generalisation is the goal: the model must perform well on data it has never seen before
Generalisation is the goal: the model must perform well on data it has never seen before

Memorising the training data is useless. The entire point of training a model is for it to handle new, unseen data. This ability to apply learned patterns to fresh inputs is called generalisation.

Where It Breaks

If the real-world data looks completely different from the training data, the model's predictions will fail
If the real-world data looks completely different from the training data, the model's predictions will fail

A model is only as good as its training data. If the data it encounters in production has shifted or contains unexpected edge cases, the model will fail silently and confidently.

The Quick Version

  • Traditional programming writes the rules; ML learns them.
  • You provide the inputs and outputs; the algorithm finds the mapping.
  • Supervised, unsupervised, and reinforcement are the three main paradigms.
  • The goal is generalisation to unseen data, not memorisation.
  • Distribution shifts and bad training data will break the model.

What to Read Next