Welcome back to Beyond the Patterns. Today I have the great pleasure to announce Gary

Marcus as one of our invited speakers. Gary Marcus is CEO and founder of Robust AI, well-known

machine learning scientist and entrepreneur, author and professor emeritus at New York

State University. Dr. Marcus attended Hampshire College where he designated his own major

cognitive science working on human reasoning. He continued on to graduate school at Massachusetts

Institute of Technology where his advisor was the experimental psychologist Steven Pinker.

He received his PhD in 1993. His books include The Algebraic Mind, Integrating Connectionism

and Cognitive Science, The Birth of the Mind, How a Tiny Number of Genes Creates the Complexities

of Human Thought, Kluge or Kluge, The Havocert Construction of the Human Mind, A New York

Times Editor's Choice and Guitar Zero which appeared on the New York Times bestseller

list. He edited the Norton Psychology Reader and was co-editor with Jeremy Freeman of The

Future of the Brain. Essays by world's leading neuroscientists which included Nobel laureates

May Britt-Moser and Edward Moser. Together with Ernie Davis he offered rebooting AI and

is well known to deconstruct myths of the AI community. In 2014 he founded Geometric

Intelligence, a machine learning company. It was acquired by Uber in 2016. In 2019 he

founded Robust AI and acts currently as Robust AI's CEO. It's a great pleasure to have him

here and I'm very much looking forward to his presentation and also a vivid discussion

with one of the main critics of pure deep learning.

Thanks very much. I'm glad to be here. This is Vancouver, British Columbia, just outside

of Vancouver, British Columbia where I am calling you from. Sorry that I'm not there

in person. And we need to do something so I can actually. Okay, good. So I think we

all know that current AI is filled with promises. Like Jeff Hinton saying in 2016, it's just

completely obvious that within five years deep learning is going to be better than radiologists

or Elon Musk estimating that in the middle of 2020 Tesla's autonomous system will be

improved to the point where drivers wouldn't have to pay attention to the road. So it's

very easy to get excited about AI. Delivery often falls short. On the left I have a chart

from Hugh Harvey. He put out on Twitter. It's a little bit old but I think it's still probably

roughly correct. The number of AI radiology startups globally was approaching 400. The

number of actual radiologists that was replaced was zero. So a lot of these promises we don't

really deliver on. It turns out that problems are much more difficult than we thought. It

turns out we don't even have enough radiologists because they get scared away because they

think they're going to be replaced. And I have a video. I won't show it right now but

of people trying to get their Teslas just to go across the street. Maybe I can show

the video. Yeah, I can't. That's all right. Just trying to get a Tesla to go across the

street without a human driver can be a pretty dangerous thing. This is the smart summon

feature. And of course the number of people have died in Teslas and so forth. So there's

been a history of over promising in AI. It actually goes back to 1950s. It's nothing

new but I think that we have to ask why. So when people couldn't do great AI in 1950s,

if they were lucky, a K of RAM, they had no data and the processing was measured in hertz

not gigahertz. But we have much more powerful technology. We still have problems. I've been

concerned about deep learning, which is the thing most people talk about all the way back

to 2012 and really back to my book, The Algebraic Mind in 2001. When it first made the news,

it was in the front page of the New York Times in 2012. And I wrote something a couple of

days later saying, hey, slow down. This is interesting. But I said that such techniques

like deep learning, well, I have enough time to read more of it. I said realistically,

deep learning is only part of the larger challenge of building intelligent machines. And I assume

everybody in this audience is familiar with deep learning. But I will remind you that

questions at any point are okay. Such techniques, I said, lack ways of representing causal relationships

such as between diseases and their symptoms and are likely to face challenges in acquiring

abstract ideas like sibling or identical to. They have no obvious ways of performing logical