Machine Learning in One Idea: Getting Better From Examples, Not Rules

Most traditional computer programs work by following a fixed set of instructions written by a programmer: if this happens, do that. Machine learning is a different approach — instead of being told exact rules for every situation, a machine learning program is shown a large number of examples and learns its own rules by finding patterns in them.

A simple way to picture it: imagine trying to teach someone to recognize handwritten digits (0 through 9) by writing out a precise rule for every possible way each digit could be drawn — every person’s handwriting is different, so this would be nearly impossible to fully specify by hand. Machine learning solves this differently: show the program thousands of handwritten digits along with the correct answer for each one, and let it gradually adjust itself until it can correctly guess the digit in a brand-new handwriting sample it has never seen.

This "learn from labeled examples" approach is called supervised learning, and it’s behind a huge share of modern AI — from spam filters (trained on examples of spam and non-spam emails) to voice assistants (trained on examples of speech paired with the correct text). There are other approaches too, like learning through trial-and-error with rewards (used for things like game-playing AI), but showing examples and learning the pattern is the most common starting point.

The obvious limitation follows directly from how it works: a machine learning system is only as good as its examples. If the training examples are biased, incomplete, or don’t represent a certain situation well, the system’s guesses in that situation will likely be unreliable too — it never learned a rule for something it was never shown examples of.