End-to-end learning (E2E)
엔드투엔드(E2E)
An approach where one AI learns the whole path, from what the sensors see to moving the steering wheel or joints, and does it all in one piece.
Overview
Older self-driving split the job into several stages. People built a part that recognizes objects, a part that predicts movement and a part that chooses the path separately, and put rules written by people between them. End-to-end does not split the stages. It trains a single neural network as a whole, so that feeding in camera video directly outputs how much to turn the steering wheel.
It learns by watching a lot of records of people driving, so it moves naturally, like a person, even in situations that are hard to write down as rules. In return, it is hard to look inside and see why it moved that way, and it is hard to find a mistake and fix just one spot. The AI that moves a robot's arms and hands is heading the same way.
An everyday example
It is the difference between memorizing the rules in a driving manual one by one to drive, and riding for a long time beside a skilled driver and learning it with your body.
Why it matters
It shows up in the news often as self-driving companies shift their technical direction this way. The view that cars drive more naturally comes with the question of how to check safety.
Also calledE2E, end-to-end