Foundations & Models
Machine Learning
Instead of giving a program every rule, you show it many examples and let it find the patterns itself. The result is not a rulebook you can look up but a statistical model that responds to new input the way it responded to what it has seen. That is why everything depends on the data: whatever is missing or skewed in the examples is missing or skewed in the model.
An example
An insurer wants to pre-sort damage claims automatically. Instead of writing rules (“if the amount is over 5,000 and it is a car, send to review”), it shows the method 200,000 old claims along with the decisions people made at the time. The result sorts well — until someone notices that it treats claims from one particular region more strictly. Not because anyone programmed that, but because more claims were rejected there in the historical data. The model learned the past, not the rule.
A closer look
A learning method searches a space of possible functions for the one that best explains the examples. This takes three things: an error function that makes wrongness measurable, a procedure that reduces the error step by step, and enough examples for the result to hold for unseen cases as well.
Why it matters
The decisive difference from classical software is troubleshooting. With a program, you read the rule and change it. With a learned model there is no rule to read — there is only data that led to a behavior. Machine learning therefore shifts the effort: away from writing the logic, toward sourcing, checking and maintaining the examples. Anyone planning an ML project who does not budget at least half of the effort for data work is planning wrong.
Commonly misunderstood
Learned does not mean understood. A model can reproduce perfectly a correlation that held during the training period and no longer does afterward. It does not notice this on its own.