learn/basics
What machine learning is
Machine learning is a way of making computers do useful work by showing them examples and letting them find the patterns, rather than by writing out the rules yourself.
Ordinary software does what a person told it to do. Someone works out the rules, writes them down as instructions, and the computer follows them. That works when the rules can be stated. Much of science cannot be stated that way. Nobody can write down, step by step, how to tell one cell type from another down a microscope, or which shape a protein will fold into. Machine learning takes a different route. You collect examples where the answer is already known, and a program adjusts itself until it gets those examples right. What it adjusts are numbers, often a great many of them, inside a mathematical function. The finished function is called a model.
The adjusting is called training. The model starts off useless, guessing at random. Each example it sees shows how wrong it was, and the numbers shift a little to make the error smaller. Repeat this across many examples and the model begins to capture the regularities in them. Nobody writes the rules it ends up using, and often nobody can read them out afterwards. The point of the exercise is what comes next: the model is given new cases it has never seen, and the hope is that the patterns it picked up carry across. This is the difference between memorising the examples and learning from them.
Why it matters for science
Scientists reach for machine learning when the pattern is real but nobody can articulate it. The evidence that a pattern exists is often just that a skilled person can see it: an expert reads the scan, hears the difference in the recording, knows the specimen on sight. Machine learning turns that judgement into something a machine can apply, by learning from cases the expert has already labelled. It also handles scale. Many fields now produce far more data than anyone can look through by hand, and a model that is merely competent can work through all of it rather than a sample.
It is worth saying how this differs from classical statistics, which also learns from data. Statistics is usually built to explain: you propose a simple relationship, fit it, and read off what each factor contributes and how sure you can be. Machine learning is usually built to predict: the model may be far too tangled to interpret, and is judged mainly by whether it gets new cases right. The two overlap and borrow from each other constantly. The difference is one of emphasis, not a wall.
Limits
A model only knows its examples. If those examples were gathered in one place, from one population, with one instrument, it may fail quietly on anything else, and it rarely announces that it is out of its depth. It will happily reproduce any bias in how the examples were collected or labelled. It finds patterns that hold, not causes, so it can latch on to something incidental that happens to track the answer. And because the learned rules are often unreadable, a wrong answer can be hard to account for.