MODULE 1 · LESSON 1
Free — no login requiredSign in to track progress, save quiz attempts and enrol in the full course.
Sign in to track progress / enrolMachine Learning in Plain Language
Here is the whole idea, and it genuinely is this simple:
Machine learning is software that learns to turn one thing into another thing, by being shown lots of examples of both.
That is it. Everything else — neural networks, deep learning, training, models — is detail about how it does that.
The difference from ordinary software
Traditional software works by rules a person wrote down. To build a spam filter the old way, you would sit down and write rules: if the subject line contains "FREE VIAGRA", mark as spam. If the sender is not in the address book and the message has six exclamation marks, mark as spam.
This works, briefly. Then spammers write "F.R.E.E V1AGRA" and your rule fails. You add another rule. They adapt. You are now maintaining ten thousand rules and losing.
Machine learning inverts the process. Instead of writing rules, you collect examples: a million emails, each one labelled by a human as spam or not spam. You hand those to a learning algorithm and it works out the patterns itself. Nobody ever tells it about exclamation marks. It notices.
The output is a model: a piece of software that takes a new email it has never seen and outputs a guess.
Every row in that table is the same shape: something in, something out. Practitioners often call it A to B. When someone proposes an AI project, the single most useful question you can ask is: what is the A, and what is the B? If nobody can answer crisply, the project is not ready.
Where the examples come from
The examples are called training data, and each one needs the right answer attached — the label. A spam filter needs emails already marked spam or not. A fracture detector needs X-rays a radiologist has already read.
This has a consequence that surprises people: the labels usually have to be created by humans, and that is often the expensive part of the project. Companies routinely spend more on getting data labelled than on the algorithm.
It has a second consequence that matters more: the model learns whatever is in the examples, including the things you did not intend to teach it. If your historical decisions were inconsistent, the model learns to be inconsistent. If they were biased, it learns the bias with perfect fidelity. Module 4 covers a real case where exactly this happened at a company with world-class engineers.
Why it works at all
It is reasonable to be sceptical. How does showing a program a million emails let it judge an email it has never seen?
Because most useful patterns repeat. Spam is written by people trying to achieve the same things — urgency, a link, a payment. Fractures look like fractures. Fraudulent transactions cluster in recognisable ways. A system that is very good at spotting statistical regularities and extending them to new cases will be right most of the time.
Note the phrase: most of the time. Machine learning systems are not correct; they are accurate to some degree. A model that is 95% accurate is wrong one time in twenty, and it does not know which time. That single fact drives most of what appears later in this course.
The A-to-B framing explains a puzzle you have probably noticed. AI reads chest X-rays at a level that impresses radiologists, and also cannot reliably tell you why your invoice is late.
Compare the two problems:
Chest X-ray → fracture present. The input is standardised: same machine, same body part, same framing. The output is a small, well-defined set. Hundreds of thousands of labelled examples exist because hospitals have been storing radiographs with radiologists' reports for decades. The relationship between input and output is entirely contained in the image.
Invoice → why it is late. The input is not one thing. The answer might live in an email nobody logged, a phone call, a customer's cash-flow problem, a dispute over a delivery, or a clerical error in a different system. There is no dataset of past invoices labelled with true causes, because nobody ever recorded the cause in a consistent field. The information needed to answer is mostly not in the data at all.
The technology is identical. The difference is entirely in the problem's shape: is the answer actually present in the input, and do labelled examples exist?
This is the most valuable diagnostic in this course, and you do not need a technical background to apply it. When a vendor demonstrates something impressive, ask what the A was, what the B was, and where the labelled examples came from. If the answers are clean, the demo is probably real. If they are vague, you are looking at a research project wearing a product's clothes.
A logistics manager wants to "use AI to reduce delivery delays". What is the first thing to establish?
Machine learning
Click to flipSoftware that learns to map an input to an output by being shown many labelled examples of both, rather than by following rules a person wrote.
Click to flip backMachine learning is software that learns an input-to-output mapping from labelled examples instead of from rules a human wrote. The practical test for any proposed AI project is to name the A and the B and say where the labelled examples come from. And remember what a model actually gives you: not correct answers, but answers that are right most of the time, with no signal about which ones are wrong.