arXiv
AI/ML
Mathematicians pin down exactly when a learner can succeed from rigged data
Imagine trying to learn what "good apples" look like when you're only ever shown good apples — the puzzle is figuring out which kinds of concepts you can still learn perfectly even without seeing bad ones.
This means we now have a complete mathematical rule for when learning systems can work with one-sided data, settling a question that's been open since the 1980s.
Bug reported: No