WHAT does “meaning” mean? It might sound like a strange question, but it has been flummoxing AI experts for decades – and philosophers for much longer. But now there’s a robot that can learn the meaning of objects and words as naturally and usefully as we do, almost like a baby in fact.
“Meaning” has been a particular bugbear for AI because you can’t build machines with human-like intelligence unless they have a notion of what objects and concepts mean. Traditionally, AI has tried to tackle this problem by working out ways for software to store symbolic “representations” of objects. For example, to store the idea of an apple, say, AI engineers specify what qualities, such as shape and colour, signify “appleness”.
But this, according to Paul Cohen at the University of Massachusetts, is precisely where we’ve been going wrong. “We are trying to make machines that acquire meaningful representations of the world with as little intervention as possible,” he says.
Advertisement
Teaming up with researchers from the University of Maryland and Imperial College London, Cohen has developed a simple robot that can move about the world and form concepts of objects around it entirely by itself, based purely on sensory information. It’s then able to use these concepts to plan its behaviour.
For example, if you hold a cup in front of its camera and start talking about the cup, the robot will begin to form an understanding of what a cup is, albeit in very basic terms. By saying “this is a cup” and “the cup is yellow” it will form simple concepts of “cup” and “yellow”, so that if you ask the robot at a later stage to turn towards the cup, or move towards something yellow, it will willingly oblige.
This is impressive, because it’s analogous to the primitive learning of a newborn baby as it first starts figuring out how to put together images and sounds.
While learning routines are everywhere in AI labs, they all involve people determining what a robot or a piece of software should learn, even with so-called unsupervised learning routines.
But Cohen’s technique puts no constraints on how the robot represents the information it acquires. Instead it uses a technique called “clustering” to find relationships between the flow of information it receives. “We don’t even tell it what a ‘word’ is, it has to figure that out for itself,” he says.
It’s a subtle distinction, but an important one, equivalent to the difference between assuming that people are born with concepts already programmed in their brains, or that they develop them through experiences. The general consensus is that the latter seems most likely since the former would be extremely inefficient and would limit what we are able to learn.
“We know that people’s memories are stored in the brain in neurons but we don’t know how they are stored at the neuron level,” says Niall Adams at Imperial College London, who collaborated on the project. We don’t suppose that concepts for objects are hard-wired into these neurons from birth.
Key to this approach is a definition of “meaning” derived by philosopher Fred Dretske, at Duke University in Durham, North Carolina. This says that for a representation to be meaningful, it must somehow have a bearing on how that person or thing acts. This is crucial because meaning comes from interacting with the environment. “You can engineer this, but it turns out to be very time-consuming,” says Cohen.
