Fei Fei Li

speaker

223 appearances 1 recordings 1 series first heard Dec 2024 last heard Dec 2024

Fei Fei Li’s voice in public audio — every appearance, attributed to the second.

Trend

recordings per month · last 12 months

No recordings in the last 12 months.Older appearances are listed below; set an alert to hear about the next one.

Appearances

newest first · ▶ plays the moment

Armchair Expert with Dax Shepard · Fei Fei Li (on a human-centered approach to AI) · 11 Dec 2024 podcast

Right. Up to that point, you can think no matter how powerful the computer was, it was used for programmed calculation. So what was the inflection concept? I think two intertwined concepts. One is reasoning. Like you said, if I ask you a question, can you reason with it? Could you deduce if a red ball is bigger than a yellow ball, a yellow ball is bigger than a blue ball.

Therefore, the red ball must be bigger than the blue ball.

Without directly saying red ball is bigger than the blue ball. So that's a reasoning. So that's one aspect. A very, very intertwined aspect of that is learning. A calculator doesn't learn whether you have a good 10 button or not. It just does what it is.

Once I had a bad one. So artificial intelligence software should be able to learn. That means if I learn to see tiger one, tiger two, tiger three, at some point when someone gives me tiger number five, I should be able to learn, oh, that's a tiger, even though that's not tiger one, two, three. Right. So that's learning.

But even before the Dartmouth workshop, there were early inklings, like Alan Turing's daring question to humanity, can you make a machine that can converse with people, QA with people, question and answer, so that you don't really know if it's a machine or a person. It's this curtain setup that he conjectured. Yeah.

So it was already there, but I think the founding fathers kind of formalized the field. Of course, what's interesting is for the first few decades, they went straight to reasoning. So they were less about learning. They were more about reasoning. They were more about using logic to deduce the red ball, yellow ball, blue ball question.

So that was one branch of computer science and AI that went on during the years, predated my birth, but during the years of my formative years, without me knowing, I wasn't in there.

But there was a parallel branch. That branch was messier. It took longer to prove to be right. But as of last week, we had the Nobel Prize awarded to that, which was the neural network. So that happened again in a very interesting way. Even in the 50s, neuroscientists were asking questions, nothing to do with AI, about how neurons work.

And again, my own field, vision, was the pioneering study about cat mammalian visual system. And Hubel and Wiesel in the 1950s and 60s were sticking electrodes into cat's visual cortex to learn about how cat neurons work. Details aside, what they have learned and confirmed was a conjecture that our brain or mammalian brain is filled with neurons that are organized hierarchically.

They're not like thrown into a salad bowl. Right. Okay. And that means information travel in a hierarchical way.

Yes. For example, light hits our retina. Our retina sends neural information back to our primary cortex. Our primary cortex processes it, sends it up to... to, say, another layer, and then it keeps going up. And as the information travels, the neurons process this information in somewhat different ways. And that hierarchical processing gets you to complex intelligent capabilities.

Or this tiger sneaking up on me.

Yes. And how did evolution assemble us so that we can recognize all this beautiful world? Not only we can recognize, we can reason with it. We can learn from it. Many scientists have used this example is that children don't have to see too many examples of a tiger to recognize a tiger. It's not like you have to show a million tigers to children. So we learn really fast.

Exactly. So just to finish, so the neuroscientists were studying the structure of the mammalian brain and how that visual information was processed. Fast forward, that study got the Nobel Prize in the 1980s because it's such a fundamental discovery. But that inspired computer scientists.

So there is a separate small group of computer scientists who are starting to build algorithms inspired by this hierarchical information processing architecture.

No, it's a whole algorithm, but you build mathematical functions that are layered.

So you can have one small function that process brightness, another that process curvature. I'm being schematic. And then you process the information. But what was really interesting of this approach is that in the early 80s, this neural network approach found a learning rule. So suddenly it unlocked how to learn this automatically without hand code. It's called backpropagation.

And also Jeff Hinton, along with others who have discovered this, was awarded the Nobel Prize last week for this. But that is the algorithm neural network.

You could actually.

You just keep filtering it. Of course, you combine it in mathematically very intricate way, but it is like layers of filtration a little bit.

Showing 121–140 of 223 · page 7 of 12