Mistrained reinforcement learning
jeudi 2 janvier 2020 à 01:00Today's AI technique of reinforcement learning has a fundamental insecurity: messing with the training data can create a back door which can be controlled by presenting a trigger in the incoming data.
Thus, a driverless car controlled by a reinforcement learning system could be mistrained to make it drive off the side of the road (which could be a cliff edge) when it sees a particular object on the other side of the road.