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

for linguistics. So WordLab had nothing to do with AI. It had nothing to do with vision. But what happened for my own North Star is that I was obsessed with the problem of making computers recognize millions of objects in the world. Why I was obsessing with it. I was not satisfied because my field was using extremely contrived data sets, like data set of four objects or 20 objects.

I was really struggling with this discrepancy because my hypothesis was that we need to learn the much more complex world. We need to solve that deeper problem than focusing on a very handful of objects. But I couldn't really wrap my head around that.

And then again, Southern California, I remember that Biederman number in my book is that I read a psychologist paper, Irv Biederman, who was up till two years ago, a professor at University of Southern California. He conjectured that humans can recognize tens of thousands of object categories. So we can recognize millions of objects, but categories are a little more abstract.

Yeah, sedan, fighter jet, and all that. So he conjectured that, but that conjecture didn't go anywhere. It was just buried in one of his papers, and I dug it out, and I was very fascinated. I called it the Biedermann number because I thought that number was meaningful, but I don't know how to translate that into anything actionable because...

As a computer scientist, we're all using data sets of 20 objects. That's it. And then I stumbled upon WordNet. What WordNet was, was a completely independent study from the world of linguistics. It was George Miller, a linguist in Princeton.

He was trying to organize taxonomy of concepts, and he feels alphabetically organized dictionary was unsatisfactory because in dictionary, an apple and an appliance would be close to each other, but that apple should be closer to a pear. Oh, I see. Then appliance. So how do you organize that? How do you regroup concepts?

So he created WordNet, which hierarchically organized concept according to meaning and similarity rather than alphabetical ordering.

That was ConvNet, Convolutional Neural Network.

So that was Young-Kung's work in Bell Labs. That was an early application of neural network in the 1980s and 1990s, where that neural network at that time was not very powerful, but But giving enough training example of digits, the scientists in Bell Labs were able to read from zero to nine or the 26 letters. And with that, they created an application to read zip codes to sort mail.

It was a lot of handwritten digits.

That data set was probably tens of thousands of examples, but we're talking about just letters and digits.

Exactly.

So I think what you were referring to was the process of making ImageNet, right?

And that process was once we realized, thanks to the inspiration of WordNet and also Biedermann's number and also many other previous inspiration, we realized what computers really need is big data. And that was so common today because everybody talks about big data, you know, OpenAI talks about big data. But back in 2006, 2006, 2007, that was not a concept.

But we decided that was the missing piece. So we need to create a big data set. How big is big? Nobody knows. My conjecture went with Biederman's number. Why don't we just map out the entire world's visual concept? Oh, my God.

Why don't we?

Okay. So Professor Kai Li at Princeton, he was very supportive of me. He was a senior faculty. But what was really critical was he recommended his student to join my lab, Jia Deng. And Jia was just a deer in the headlight as a young first-year graduate student. He didn't know what's going on. He got this crazy assistant professor of me.

and told him that we're going to create a data set that map out the whole world's visual concept. He's like, sure. You know, I don't know what you're talking about, but let's get started. So he and I went through the journey together. I mean, he's a phenomenal computer scientist and many hoops we jumped through together. It was just the solution that got us through.

That's an interesting observation.

I do tell my kids ideas are cheap. Exactly. Hollywood.

Showing 141–160 of 223 · page 8 of 12