How a Machine Actually Learns

A three-year-old can identify a dog. Not one specific dog, any dog. A Labrador, a street dog in Chattogram, a cartoon dog, a dog seen from behind in poor light.

Nobody gave the child a definition. Nobody listed the criteria. They saw perhaps forty dogs over two years, someone said “dog” each time, and something in their head worked out the rest.

Machine learning is an attempt to do the same thing with a computer. That is genuinely the whole idea. Everything technical about it is detail on top of that one move: stop writing the rules, show examples instead, let the system work the rules out.

What “learning” means mechanically

The word learning makes this sound mysterious. It is not. Underneath, a model is a large collection of numbers, and learning means adjusting those numbers until the output stops being wrong.

The loop has four steps and it never changes.

1. Guess. The model looks at an example and produces an answer. At the start these numbers are random, so the first guess is nonsense.
2. Check. Compare the guess to the correct answer, which you already have because this is a training example.
3. Measure the error. Not just right or wrong. How wrong, and in which direction.
4. Adjust. Nudge the numbers slightly so that the same input would produce a slightly better answer next time.

Then repeat. Millions of times, across thousands of examples.

No single adjustment matters. The first thousand rounds produce a model that is still terrible. But the errors push consistently in a direction, and after enough repetitions the numbers settle somewhere that works. Nobody decided what those numbers should be. They were found by being wrong repeatedly in a structured way.

This is why machine learning needs so much data and so much computing power. It is not clever. It is patient. The intelligence is in the process, not in any moment of it.

Three ways to set up the problem

Supervised learning

You have examples and you have the right answers. Ten thousand emails already labelled spam or not spam. Photographs of leaves labelled diseased or healthy. Loan applications labelled repaid or defaulted.

The model learns to reproduce the labels, then applies that to cases it has never seen. This is the most common kind by a wide margin, and the one behind almost every practical application you have used.

Its weakness is obvious the moment you try it. Someone has to produce the labels. For ten thousand medical scans, that someone is a radiologist working for weeks. The bottleneck in most real projects is not the algorithm, it is the cost of the answers.

Unsupervised learning

You have data and no answers, and you ask the model to find structure in it.

A shop with fifty thousand customers and no idea how they differ can ask a model to group them by buying behaviour. It might return five clusters. The model does not know what they are, since nobody told it anything. A human then looks and recognises bulk monthly shoppers, evening top-up buyers, and so on.

This finds patterns nobody thought to look for. It also produces groupings that are meaningless as often as not, and only a human who understands the domain can tell which is which.

Reinforcement learning

No examples at all. The model acts, receives a reward or a penalty, and adjusts to earn more reward.

This is how systems learn to play games far better than people. It is closer to how a child learns to walk than to how they learn vocabulary. It works beautifully where you can run millions of cheap attempts, which is why games are the showcase and why applying it to a physical robot, where each failed attempt breaks something, remains hard.

The failure that catches everyone

There is one mistake so central that understanding it is most of what separates people who can do this work from people who have completed a course.

I see the human version of it every exam season. A student memorises five years of past papers. They score well on those five papers. Then the real exam asks the same concepts in a different form, and they have nothing, because they learned the papers rather than the subject.

Models do exactly this, and it is called overfitting. Given enough capacity, a model will memorise its training data rather than learn the pattern underneath it. Performance on the training set looks superb. Performance on anything new collapses.

The defence is simple to state. Hold back a portion of your data before training and never let the model see it. Judge the model only on that held-back portion. If it scores 98 percent on data it trained on and 61 percent on data it did not, you have not built a detector. You have built an expensive memory.

The reason this matters beyond the technical detail is that overfitting is invisible from inside. Nothing warns you. The numbers look better than ever, right up until deployment.

Where it goes wrong in the world

The original version of this post mentioned bias, privacy and scale as challenges and left it there. They deserve more than a list.

Bias is inherited, not invented

A model has no values and no opinions. It reproduces whatever pattern exists in the data it was given.

Train a hiring model on ten years of a company’s decisions and it learns those decisions, including the ones nobody would defend out loud. If the company rarely promoted women into technical roles, the model learns that pattern as reliably as it learns anything else, and it will apply it with perfect consistency and no explanation.

This is not the model malfunctioning. It is the model working exactly as designed on data that encoded a problem. Removing the obvious field does not fix it either, because other fields quietly carry the same information. The failure sits in the data and in the decision to automate without examining it.

Privacy is a design choice made early

Models need data, and useful data is often personal. Medical records, transaction histories, messages, locations.

There are real technical answers now. Federated learning, one of the areas I work in, trains a model across many devices without the raw data ever leaving them. Only the learned adjustments travel back, not the records. It is not a complete solution, and it is a serious improvement on collecting everything into one place and hoping.

A model cannot explain itself, and that is sometimes unacceptable

The most accurate models are the hardest to interrogate. Ask why a particular loan was refused and the honest answer involves millions of numbers with no individual meaning.

For a film recommendation this does not matter. For a loan, a medical diagnosis or a criminal risk assessment, “the system said no” is not an answer a person can appeal or a regulator can accept. Explainable AI is the research area addressing this, and it is one of my own, because a model that cannot justify its output is not deployable in the places where the stakes are highest.

What it cannot do

Two limits are worth holding onto, because most overselling of this technology depends on people not knowing them.

A model finds correlation, not cause. If ice cream sales and drowning deaths rise together, a model will happily use one to predict the other. It has no concept that summer causes both. This is fine for prediction and dangerous the moment anyone treats the output as an explanation of why something happens.

A model fails on what it has not seen. Trained on Chattogram traffic, it will not handle Dhaka. Trained on transactions from before a new fraud technique existed, it cannot recognise that technique. Models are confident on unfamiliar input, not cautious, which is precisely the wrong instinct and the reason monitoring after deployment matters as much as testing before it.

Why I work on this

My research applies these methods to phishing and smishing detection, to medical imaging, and to identifying crop disease from photographs. The last one is the clearest case for why any of it matters here.

A farmer photographs a diseased leaf and gets an identification in seconds. Doing that well requires a model trained on the crops actually grown in Bangladesh, running on an ordinary phone with an unreliable connection. It is not the most sophisticated problem in the field. It is one where getting it right changes a harvest, and the technical requirements are shaped entirely by the conditions rather than by what looks impressive in a paper.

That is the part of this work I find worth doing. Not the systems that learn something remarkable, but the ones that learn something ordinary and then survive contact with the place they are meant to be used.

---

I research machine learning, quantum computing and AI-driven cybersecurity, and teach Computer Science at Sunshine Grammar School and College in Chattogram. My publications are on the research page.