Great on familiar work, bad on new work
Last time, AI traced its mistakes back to turn the right knobs. But turning them too carefully can hurt: the machine can memorize the test without understanding the lesson.
EXAMPLE
Two students studying for a test
Picture two students studying for a math test. An memorizes the answers to 100 practice problems. Binh learns how to solve them, even though he still gets a few wrong.
NEW TEST
Turn over the test
On test day, only the numbers change. An freezes, because he never memorized this one. Binh can still do it, because he knows the method, not the answers.
MEMORIZING
Machines memorize too
Machines do the same. Turn the knobs too carefully to fit every training example, and it memorizes even tiny details. That's called overfitting: nearly perfect on practice, wrong on new data.
HIDDEN TEST
Always keep an unseen test
That's why trainers always set aside part of the data that the machine never sees while learning. Only after learning is it tested on that. The score on unseen questions is the real score.
WHY
Familiar is smooth, new is a guess
Chat AI learns from a huge amount of text, so it answers common questions very smoothly. In a new situation nobody has written about, it may guess wrong and still sound confident.
PART 6
Why did AI suddenly get so good?
Learning to understand is good. But why has AI suddenly gotten so much better in the last few years? See you in part 6.
This article is based on the video Memorizing vs. understanding from the Mark học AI channel. Watch the video (in Vietnamese) to see the animations.