Can AI keep learning from experience as humans do?
Serious leads
Why it matters
Neural networks tend to forget earlier tasks when they learn new ones. Researchers call learning tasks one after another crucial to developing artificial intelligence: solving it could let AI systems keep acquiring new skills over time without losing old ones.
What we know
Today's models mostly stop learning once training ends.
What blocks us
Learning new things tends to erase old ones (catastrophic forgetting).
Current leads
External memory systems, targeted and sparse updates of the model.
Sub-questions
Why is it still so hard for an AI to learn a new kind of data without forgetting what it learned before?
Can an AI keep learning without storing and replaying old examples?
How can learning a new task make an AI better at earlier tasks, instead of worse?