Radical Reads

AI pioneer, Vector Institute Chief Scientific Advisor and Turing Award winner Geoffrey Hinton published a paper last week on how recent advances in deep learning might be combined to build an AI system that better reflects how human vision works. Hinton’s system is called “GLOM” and in this exclusive Q&A with Radical partner Aaron Brindle, Geoffrey explains how it works, its implications for everything from self-driving cars to natural language processing, and why he landed on the term (or acronym?) GLOM.

Aaron Brindle: What’s the biggest difference in how a human brain processes an image versus current neural networks?

Geoff Hinton: For starters, there are enormous differences in the hardware. The brain processes images using a huge number of connections at low power. Computers have fewer connections but loads more power. Computer vision models, historically, have looked at single images where a static picture is presented at a uniform resolution. Traditional AI vision systems try to process the entirety of that uniform image.

That’s completely different from what people do. For humans, vision is really a sampling process, where the eye makes real time decisions around what information in the field of vision is going to be further deciphered. For example, we’re very good at quickly sampling and processing anything that moves. The same is true for something that has a different colour from its background. When the eye fixates, whatever is in the middle of the retina is at high resolution and whatever is around the edge of that will be at a low resolution. You process what’s in focus for several 100 milliseconds before your eye fixates on something else. GLOM is about the deep learning process that happens after the system fixates on an image. It addresses a research problem I’ve had for the past 50 years and feels much closer to an understanding of how the brain might be doing vision.

AB: So what is GLOM solving for?

GH: Deep learning has been good for things like perception and for motor control and for manipulating objects. There’s a belief among some people in the research community, however, that deep learning is not so good at tackling symbolic representations. I think they’re wrong. And this paper is about how to do one aspect of the symbolic stuff that people thought couldn’t be done easily with neural networks. In the paper, I discuss how islands of identical vectors can be used to represent the part-whole hierarchy.

AB: Is the objective here a more accurate system?

GH: Hopefully it makes for more accurate systems, but also more interpretable systems that work more like the way people do. If you can get GLOM to work, you may be able to get rid of things like adversarial examples.  Adversarial examples show that traditional convolutional networks are currently doing it completely wrong. With GLOM, you should be able to recognize images based on the relationships between parts of an image, not on fine textures and things that aren’t discernible to the human eye.

AB: If GLOM works, how will it change computer vision systems?

GH: I think vision systems will behave more like people, which is particularly important for self-driving cars. But a system like this could be applied everywhere: from satellite imagery that can determine how agricultural land is being used, to an up-close analysis of a leaf that can determine the health of an individual plant. It will also be very useful for language, as this same system when applied to natural language models may make it much easier to interpret how the language system is understanding phrases and sentences.

AB: I notice that your references to GLOM are in all caps. What does GLOM mean and is there an acronym you haven’t shared?

GH: I landed on the term, originally, from the slang for agglomeration – the idea of things that “glom together.” But it might also stand for “Geoff’s Last Original Model.”

AB: Let’s hope not!

GH: We’ll see.