Machine learning is what you reach for when a rule exists but nobody can write it down. Ordinary software carries out instructions somebody typed; a learned system is shown inputs alongside the answers you wanted and derives the mapping itself. Nobody can specify line by line what makes a photograph contain a bicycle, or which invoices will go unpaid, yet both leave traces across enough cases for a model to pick up. The whole craft compresses into one word, generalisation: being right about situations that were absent from training. A model that reproduces its training set flawlessly and stumbles on anything fresh has failed, and it has failed in the most flattering way available, which is exactly why this discipline is built around withholding data and testing against it.
Readers ask constantly what divides this from statistics, and it is not, disappointingly, the mathematics. Regression, likelihood, optimisation and the trade-off between bias and variance belong to both. What differs is the question and the referee. A statistician fits a model to say something defensible about a population: which effect is real, how large, how certain, under which assumptions, with coefficients a human being has to interpret. A practitioner here fits a model to minimise error on data kept back from training, and will happily accept something nobody can interpret provided it wins on that data. Hence one side worries about violated assumptions while the other worries about the test set leaking into training. Neither party is the serious one. They answer to different judges, and plenty of confused hiring follows from pretending otherwise.
Three ways a model gets its lessons
Supervised learning is the workhorse: every example arrives with the right answer attached, a spam flag or an eventual sale price, and the model learns the route from one to the other. Unsupervised learning gets no answers and hunts for structure instead, grouping customers whose behaviour rhymes or isolating the transaction that resembles nothing around it. Reinforcement learning has no fixed answers either, only a score it earns by acting, which suits games, routing and control. In practice the borders blur and most production work is supervised, simply because most organisations already keep records of what happened next. Neural networks and deep learning cut across all three; they are a family of models within the field rather than an alternative to it. Every style rests on data analysis done beforehand, because a model inherits whatever is wrong with the table behind it, and inherits it without comment.