Artificial intelligence has passed a classic theory of mind test used with chimpanzees. The test probes the ability to perceive the world from the view of another individual and so AIs with this skill could be better at cooperating and communicating with humans and each other.
AIs with theory of mind are key to building machines that can understand the world around them. In recent years, the skill has emerged in a robot whose memories are modelled on human brains and in DeepMind’s ToM-net, which understands that others can have false beliefs.
Advertisement
Raul Vicente at the University of Tartu in Estonia and colleagues drew inspiration from an animal study that looked at the feeding habits of dominant and subordinate chimpanzees. “Chimpanzees have these hierarchical structures in their society and in principle the dominant one almost always gets the food,” says Vicente.
The chimp study showed that the subordinate animal would only go for food that it knew the dominant animal couldn’t see it, suggesting an ability to place itself in another’s position.
To replicate this set-up, Vicente and colleagues created a virtual 11 by 11 grid in which they placed two AIs – one dominant and one subordinate – and a single piece of food in different orientations and locations.
The subordinate AI was able to move within the grid, and was rewarded points if it ate food that the dominant agent couldn’t see, but lost points if it ate food in the dominant agent’s sight.
It learned via trial and error – in a process called reinforcement learning – whether or not to move towards the food.
Apes and AI
A key difference between apes and AIs is that the AI required several thousand trials before it learned the task, while chimpanzees understood it intuitively.
Being able to understand the perspective of other individuals enables communication, cooperation and competition, says Vicente. “It is absolutely essential for survival for any social species.”
The ultimate goal of this kind of research is for machines to become effective at communicating with people, but it will be a long time before AIs develop theory of mind at a level comparable to humans, says Vicente.
Eventually, the goal is to develop AI that can understand what we think and see. “It will be fun and effective to interact with these agents – when they are able to put themselves in our perspective,” says Vicente, because AIs with theory of mind would be able to communicate things that we don’t know or can’t see.
But Joanna Bryson of the University of Bath in the UK says that AIs that develop the ability to see from other perspectives won’t necessarily lead to machines that are more like humans. Theory of mind in humans is magnified by the fact that we have language to express our thoughts and feelings, she says.
Complex social interactions that involve verbal communication and body language can’t yet be reflected well in virtual tests.
The next stage of the research is to see whether AI understanding of perception holds up in different and more complex environments.
Reference: arXiv,Â
