Artificial intelligence: machines should learn how to perceive people

For a good half a century, Geoffrey Hinton has been thinking about how the brain works and how you can model its circuits in a computer. So far he has been quite successful with it: in 1986 he was largely responsible for spreading the idea of ​​”backpropagation” – a process with which artificial neural networks can be trained efficiently. And in 2012, together with colleagues, Hinton was able to show for the first time that a deep neural network can recognize images better than a person – the work triggered a rapid development in deep learning.

In November last year, now 75, he had a new idea. “It’s my best bet right now on how things will fit together,” says Hinton, who has sealed himself off from the pandemic in his Toronto office. If this bet works, it could spawn the next generation of artificial neural networks. His “honest motivation”, as he calls it, is curiosity. But the practical motivation – and ideally the consequence – would be a more reliable and trustworthy AI that would be a lot more powerful than the best systems out there today.

Because despite the rapid advances in this area, there are still major challenges. If a neural network is exposed to an unfamiliar data set or an unfamiliar environment, it turns out to be brittle and inflexible. Self-driving cars and speech generators that write essays are impressive, but a lot also goes wrong. Visual AI systems can be easily confused: a coffee cup that is recognized from the side is an unknown object to the system when viewed from above unless it has been trained to this view; and with the manipulation of a few pixels you can trick an AI into mistaking a panda for an ostrich or even a school bus.