Geoffrey Hinton
(Photo by Daniel Ehrenworth)
Presentations by deep learning pioneer Geoffrey Hinton and University of Toronto alumnus and entrepreneur Allen Lau were among the highlights of the opening day at this year’s influential Collision technology conference.
Hinton, a U of T University Professor Emeritus who now works at Google and has been dubbed the “godfather of deep learning,” spoke on Tuesday about his efforts to develop a new type of artificial neural network that can recognize and interpret images better than existing systems.
One of more than 600 presenters at the annual conference that runs until April 22, Hinton noted that most current systems use convolutional neural nets that don’t recognize objects the way people do.
“They use lots of texture information, which people are insensitive to, and they use a lot of shape information that people are sensitive to,” said Hinton, an A.M. Turing Award-winner who is also the chief scientific adviser at the Vector Institute for Artificial Intelligence.
“They also aren’t very good at extrapolating to new viewpoints. People can see an object from one viewpoint and then extrapolate to many different viewpoints. But convolutional neural nets need to see objects from lots of different viewpoints to understand them.”
To illustrate this flaw, Hinton showed images of a school bus that had been mildly obscured with visual noise. While humans can still easily recognize the images as showing school buses, Hinton revealed that convolutional neural nets mistake the buses for ostriches.
His solution? A system called GLOM – short for “agglomeration” – that’s inspired, in part, by the relationship between cells, organs and DNA.
“What I want to do is design neural nets so they have different ways of seeing the same thing, just as people do, and they see things the way people see things,” he said. “That would make them far more interpretable and far less likely to make crazy errors – like that school bus.”
For the second year running, Collision – which features global speakers from technology, business and media – is being held virtually due to the COVID-19 pandemic. The event, which counts U of T as an official partner, was slated to be hosted in Toronto for three years starting in 2019.