WHY THIS MATTERS IN BRIEF
- The world’s leading AI researcher believes that teaching AI’s common sense will lead to a revolution in how they learn and take them to the next level
Around five years ago artificial intelligence (AI) made a giant leap forward and suddenly, almost overnight, became much more accurate in interpreting images, and that led to a new revolution biometrics, such as facial recognition, and the way we search images.
The reason for the leap? The development of new artificial neural networks, and one of the world’s top AI researchers, Yann LeCun, the director of Facebook’s AI research group, recently went on record to suggest that, while there’s still progress to be made, he believes we can use these same image recognition systems to teach AI common sense, which in turn could kick start a whole new AI revolution.
Today’s machine vision systems are advanced – way more advanced than they were just five years ago – to the point that if you show them any photo that has a dominant object in them then nowadays they’ll have no trouble whatsoever identifying it – even down to the brand, or in the case of animals, the breed. They can even recognise more abstract images such as landscapes, weddings, birthday parties and sunsets – something that five years ago was thought of as being impossible.
How things change… But that said today most machine vision systems still struggle trying to caption images – such as “A dog running on a beach at sunset,” and if you show them images of other types of objects, or unusual situations they’ve never seen before then it’s likely that you’ll only get one result. Garbage. And apparently, according to Lecun that’s because machine vision and recognition systems don’t have any common sense. So what’s the link between images and common sense you ask?
Apparently one of the problems that holds back the development of AI, in LeCun’s opinion, is its lack of experiences – take humans for example. We amass knowledge over a long period of time and our brains can stitch it all together, applying the experience from one situation to another, and so on.
At the moment AI’s are trained is what Lecun describes as more of a classical way – we tell them what something is, and what you can do with it, or to it. For example, “this is an orange,” “this is a car,” “you can push this to move it.” In many respects this is similar to how we educate babies.
One of the core projects that LeCun is focusing on now however is try to get AI’s to learn common sense just by watching videos. In this kind of example an AI might see someone pushing a car and then it’s hoped that it will make the connection that a car can be pushed – and of course, all this will be able to be scaled up almost exponentially.
“You learn a ridiculously huge amount of stuff just by observation,” LeCun says, “but today there are lots of ways we can fool machines because they have a very narrow knowledge of the world.”
Now LeCun and his team are now trying to build learning systems that can predict the future – they want to show them a few frames of video and get them to try to predict what happens next. They think that if they can train a system to do that, to “learn” common sense, then they’ll have developed an entirely new learning technique that could one day support a revolutionary new unsupervised learning system. And that, again, as though we haven’t already had enough AI revolutions, would be another huge step forwards for AI.