Researchers present safe way to disable machine learning

A researcher from Google’s DeepMind team, in collaboration with a scientist from the University of Oxford, presented a study into the possibility of safely turning off a machine-learning system. They refer to a kind of ‘big red button’.

In the paper, researchers Laurent Orseau and Stuart Armstrong talk about the fact that a system that learns through reinforcement learning will not always exhibit optimal behavior. Therefore, there should be an option to disable such a machine learning system in the event that it exhibits behavior that could be harmful to the environment or to the person interacting with the system.

The challenge of the research is to find a way to disable the system without finding a way to disable ‘the big red button’ itself to ensure that it cannot be disabled. This could happen, for example, if the system expects certain rewards from harmful actions, the researchers write. They therefore present a way that prevents systems from learning this behavior or looking for ways to avoid being disabled.

This breaks down into several parts, the first part being to ensure that the human interruptions do not give the impression of being part of the task to be performed. By forcibly adjusting the system’s behavior, it should be given the impression that it has chosen a new policy itself. The researchers also come to the conclusion that certain learning algorithms are easy to interrupt, while others require some adjustments.

The part of the University of Oxford involved in the research is the Future of Humanity Institute, which, according to Business Insider, focuses on researching threats to humanity, among other things. The institute’s leader, Nick Bostrom, is convinced that artificial intelligence will be smarter than humans within a hundred years and that it may turn against humanity.

This concern is also present in a committee of the European Parliament, such as wrote Political Thursday. The site has obtained a report citing the possibility that artificial intelligence is surpassing the human variant. Research should therefore be carried out that examines how this development should be taken into account, for example when it comes to employment.

But safety is also one of the topics, as is the proposal to grant the most advanced robots a separate legal status, Politico writes on the basis of the report. These robots could be given the status of an ‘electronic person’, along with certain rights and obligations.