Science Friday - An AI Leader’s Human-Centered Approach To Artificial Intelligence

Episode Date: December 4, 2023

Just about every day there’s a new headline about artificial intelligence. OpenAI Founder and CEO Sam Altman was forced out, and then dramatically returned to his post—all in the span of a week. Then there’s the recent speculation about a revolutionary new model from the company, called Q*, which can solve basic math problems.Beyond the inner workings of AI’s most high profile startup are stories about AI upending just about every part of society—healthcare, entertainment, the military, and the arts. AI is even being touted as a way to help solve the climate crisis.How did we get to this moment? And how worried or excited should we be about the future of AI? No matter how it all shakes out, AI leader and early innovator Dr. Fei-Fei Li argues that humans should be at the center of the conversation and the technology itself.Ira talks with Dr. Fei-Fei Li, founding director of the Institute for Human-Centered AI at Stanford University and author of the book The Worlds I See: Curiosity, Exploration and Discovery At The Dawn of AI, about her path from physics to computer science and the promise and potential of human-centered artificial intelligence.To stay updated on all things science, sign up for Science Friday’s newsletters. Transcripts for each segment will be available after the show airs on sciencefriday.com. Subscribe to this podcast. Plus, to stay updated on all things science, sign up for Science Friday's newsletters.

Dr. Fay-Fay-Lee is a trailblazer in the world of AI. But how does she keep going when things don't go as planned? Once I identify that audacious quest, it is relatively easy for me to tune out the other voices. It's Monday, December 4th, and you guessed it, it's still Science Friday. I'm SciFri producer Shishana Buxbaum. Just about every day, there's a new headline about it. AI. It feels sometimes like it's about to upend every part of society. But how exactly do we get to this moment and how worried or excited should we be about the future of AI? Ira recently sat down

with Dr. Fay-Fei Lee to discuss how she came to understand the importance of what she calls human-centered artificial intelligence. Dr. Fay-Fei Lee is the author of the book, The World's I See, Curiosity, Exploration, and Discovery at the dawn of AI. She's also the founding director of the Human-Centered AI Institute that's at Stanford University out there in California. Dr. Lee, welcome to Science Friday. Thank you, Ira, for inviting me. Nice to have you. Now, I mentioned this in the – when you say that AI needs to be human-centered, explain that for us, please.

Well, okay, well, let's begin with what AI is. AI is a piece of tool. I know it's a very intricate and intriguing piece of tool that humans have made, and I do believe this is a piece of tool that is very powerful and will transform human society, will transform business and all that. But at the end of the day, it's a piece of tool. Tools are made by humans, are being deployed by humans, and should be used properly by humans. and no matter how you think about it, I put human in the very center of this technology because of our responsibility in the creation and application of it. But people view these more than just tools.

They view them as intelligent tools and fearful that they may become more intelligent than the tool makers. I hear you. I do hear you. And a lot of this is because it's new, it's unknown. and when we face something new and unknown, it's scary. And this is not the first time humanity has faced that. We think about the history when we first discover fire as a species, when we created electricity, when we created, you know, PC.

Just every step along the way, major technological advancement in human history creates anxiety and disruption. this is the same. And of course, it is an intelligent piece of tool in the sense of it takes data, understands patterns, help to make decisions, and all that. But as far as a piece of software, this is still very much a piece of tool. So you do not believe that AI poses an existential threat, correct? Let me be more nuanced with my answer. First of all, as a scholar, I do respect discussions about this. You know, look, I live on Stanford campus and where my colleagues discuss about the archaeology of Roman Empire all the way to, you know, the smallest bacteria we can find in

human bodies. So there's a lot of curiosity and where intelligent machines are going as a potential piece of software is a worthy topic. But as of now, I see AI. more urgent and pressing risks in social domain, such as disinformation for democracy, job changes, bias and privacy infringement, and then many more. It's something you can actually feel. It's something that impact everybody, impact everyday people. Let's worry more about those than about what AI is up to. I think we need to take responsibility in those issues immediately, and that's

important and because of the imbalance of the discussion of this kind of risk and issues versus the existential crisis, I feel it's my responsibility and especially being the co-director of Stanford Human Center AI Institute, we should be communicating this. Your book is part memoir, part history of AI, but I found it interesting that you had to be convinced to include your personal story in the book. right? Tell me about that. Yes. So I was invited to write an AI science book to the popular audience

about three and half years ago. I remember it was beginning of COVID. And of course, as scientists, I wrote for a year, first draft of a science book. And I showed it to my very good friend, Professor John H.Mendi, a philosopher and co-director of Stanford H.A.I. And he literally said, you have to rewrite. And it was pretty hilarious, but to me it wasn't that funny when someone told me to me to rewrite after a whole year. But he was very convincing. He said, look, Fefei, there are many

AI technologists who can focus on a pure science book, but if you are talking to the greater audience, there are so many immigrants, young women, people of all walks of life, people of all disciplines, they are lacking a voice they can identify with. And he believed that I could be embody that voice. And I think he's right. So I had to rewrite the book in the double helix structure where I use my personal journey of coming of age as a scientist to carry the very serendipitously intertwined story of AI coming of age. Wow, that's a great metaphor. I've never heard of the double helix metaphor used in telling a story, but it certainly fits. Yeah, I'm a nerd. You talk about moving from China to suburban New Jersey as a teenager. And how did this experience

shape your curiosity and eventual career in AI? Yeah, Ira, that's a great question. You know, it did dawn on me while I was writing the book. There's so much similarity far more than I thought between being an immigrant, especially learning a new language and getting to know a new country, and being a scientist. Both really propels you or put you in a situation of unknown. And then you have to explore. You have to find your inner North Star and just have that grit and determination and develop resourcefulness to go after something that you're curious. about. So in a way, maybe the immigrant experience did shape me as a scientist in the sense of be very curious and not afraid of an unknown situation. You studied physics as an undergrad.

You like physics. I loved physics. Actually, that was my first North Star. You know, between Einstein and everything. This is why I went to Princeton and majored in physics. I share this love for you. I never had the kind of intelligence to do what you do. But how did physics lead you to computer science and then on to AI? What's the connection there? Yeah. Well, when I was a kid, a teenager, kid, a rather lonely one since I didn't speak much of that. That I can relate to. Yeah, I didn't speak much of the language, was busy, you know, trying to make a living. I read a lot of Einstein and physics. I love the classes, physics classes. So I went to major in physics. In my book, I also talk about Neil deGrasse's class, right? He taught me astrophysics.

He did. He did. Great teacher. Yeah, well, amazing teacher. I did not realize who he was when he was my professor. But what I really loved about physics is the audacity to ask the most fundamental questions about the universe. The physicists are not afraid, you know, you see the stars. You see the stars. moving, and then you start to imagine a gravitational force that can be captured in one equation that explains the movement of all the heavenly bodies. You know, you go after what is the beginning of space and time. You go after questions like, what's the smallest matter? Can you break down an atom? I mean, these are just, you know, way crazy questions to ask. Yet physics as a discipline gives you both the rigor as well as the fearless curiosity to chase these questions. And that was what

I loved about it. And then in the middle of my physics study during Princeton years, I started reading great physicists of 20th century. And towards the second half of their career, they start to ponder questions beyond the physical world, like Schrodinger, what is life, Roger Penrose wrote about mind, and Einstein is always such a fluid mind of pondering about so many things, it kind of took me in an unexpected turn to become more curious about life. And once I became more curious of life, I was naturally drawn to the most mysterious, audacious question I could ask as a student at that time. which is what is intelligence?

You know, what makes me humans intelligent and can we make machines intelligent? And that led me to artificial intelligence as well as human neuroscience. So I do have, in a way, a relatively untraditional path into computer science. It was not video games, and it was not just hacking software.

it was physics. And being fearless about asking the questions. Yes. Right. Yeah. That's important in science, isn't it? Oh, it's essential in science. And believing that you can understand what intelligence is,

if you're going to make an artificial intelligence, you have to have some sort of belief that you can decipher what intelligence is. Do you not? Yes and no. I think I believe that journey. I believe that we need to go on. that quest. But what is really curious is that the process of making an intelligent machine and the process of understanding brain, human brain, is simultaneously parallel and intertwined.

The understanding of the brain inspires AI, but it's not limiting us to make a different kind of machine, thinking machine. Let's talk about neural networks. We've heard about those things. work with neural networks, what is a neural network? And how does it compare to what's going on in my own brain? Yeah, well, let's start with the most organic and amazing neural network nature has made, which is the brain. What does our brain look like? There's a piece of work that eventually won a bell price in medicine, which is by neurophysiologist, Huberon, Vizzo in the late 50s, they were wondering about how mammals see. And we really, other than knowing the functions of retina and eyes, which is really a sensor

that collects lights and send electric signal back into the brain, we don't really know how you go from photons stimulating your retina to, oh, I see a fish. That is a computational question. And they were probing a mammalial brain using electrodes. At that time, it was a very, very advanced experimental technology. But what they find out are two remarkable things about the mammalian visual brain, which eventually inspired the computer neural network. The one thing they find out, well, we know the brain is made of small,

cells called neurons. What they find out is every neuron in the cat visual brain, especially close to the retina, like in what we call early stage visual brain, it responds to something simple. It respond to, say, a moving bar that's oriented. Shades, if I remember correctly. Right, but it's really simple shape. It's really an edge, an edge of a particular or a say 45 degree to the left, moving, right, to the left. And they found out that there's millions and millions of these neurons that all respond to something slightly different at the beginning, just edges, slightly different orientation.

And then you go to the next layer where these neurons send their signal to, and this next layer respond to something slightly more complex, maybe just a corner. you know, and then you keep going. There's a hierarchy of information propagation that eventually you go high enough in the brain. There's something that correspond to, I see an object that's a fish, right? So what they found out is that the fundamental unit, computing unit of the brain, which is a neuron, respond to simple signals, and the meaning of them stacked together in the network, can give you more complex computation like seeing a fish.

And that concept, that two concepts, individual neuron units, put together in a hierarchical network that propagates information and learn about the input signal is the foundation of a neural network. This is Science Friday from WNYC Studios. If you're just joining us, I'm talking with computer scientist Fei-Fei Lee about her new book, The World's I See, Curiosity, Exploration, and Discovery at the dawn of AI. So you have to train the computer? Absolutely, you have to train.

What does that look like? You give the computer, for example, you want to train the computer to see a cup. Right. That's in front of me. You give the computer many, many, many cups in different angle, different lighting. And then the neural network has all these neurons. There are small, tiny mathematical equations. They're connected by tiny mathematical functions.

But mathematical function has parameters, right? Like you have to tune them. And then you use this training algorithm to tune these parameters. And there is a goal for this algorithm. There are different types of goals. We call them mathematically objectives. Let's just make an example. The objective here is to see this as a cup versus some

else that's not a cup. So it's a simple goal of cup, non-cup. Well, every time you give it a training picture of a cup, you tweak your parameter so that it tries to answer this picture as a cup. And if it's wrong, the system sends a signal say you're wrong and then you tweak again. And then you do this many, many, many times. You train it with cup picture or non-cup picture. and then you eventually learn. That's just one type of learning. I'm simplifying. Well, that's good, because I understood that.

I'm glad. I know you're the creator of ImageNet, which uses this algorithm we've just been talking about, right? The project wasn't exactly smooth sailing all the time, was it not met with immediate adoration? When did you realize that it would shape the field of AI so profoundly that you were right about what you were doing? Well, there's different ways of realizing you're right.

When I hypothesize this project, I was driven by the scientific mission and quest. I know that we need to use data to drive AI algorithm. So from that point of view, I was delusionally confident. I was right. I did not care that there are so many people who told me that was wrong. So that was the one way of feeling I was right. But it doesn't mean it was easy. I was facing a lot of pushbacks.

And then, of course, the project proceeded. We finished the project. And then fast forward six years later or five years later after the onset of the project, we got to the moment that the world knows as the beginning of deep learning revolution when ImageNet, Convolutional Neuronetwork, And literally two GPUs showed progress in visual, intelligent task that was really unexpectedly big. That was the moment of external validation. How do you keep going when so many people are telling you, well, maybe this is not right.

And what is there about your personality? Did you have this growing up? Well, this, this goes back to what we talked about. I, whether you call it personality or whatever, somehow I start with that North Star. As a scientist, I'm driven by North Star, that audacious quest. Right. And once I identify that audacious quest, it is relatively easy for me to tune out the other voices.

I know one of the big challenges is the bias baked into some of these algorithms that were talking about. The algorithms are only as good as the data they're based on, right, which replicates things like racism and sexism in the real world. How do we get better at that? Yeah, it's an important issue. I mean, algorithm bias is one of the main risks that AI technology brings. And there are multiple ways to mitigate this. There's the technological way I'll get into, but there is also the social norm and regulatory framework, which is also important. On the technology side, we know a lot more today now about where bias comes in. It starts with, you know, the way we design and curate data.

It has to do with the algorithm itself and also has to do with how we use the output of the algorithm. And because we now know so much more, there are technological, solutions, right? Be careful with your training data, how to balance the data. But there's also the social piece, right? Whether you're a researcher or you're developing a product, there's more and more awareness in a social context of the harm of data bias and algorithm bias, and we try to mitigate that through that. Eventually, we will need some guardrails, depending on the vertical space, whether it's health care or finance, some of the guardrails needs to be assessing and evaluating issues like bias. There are so many fits and starts in history of AI. It seems like a

game changer than a sort of fall to the wayside. Is the current AI any different than that? Are we on the right path? Is there a right path? Is that the wrong question to ask? That's a great question to ask, and it's not a wrong question. It's just, it requires nuanced answer. Let me first share with you, I do believe we're at the inflection point. I know there has been bubbles and bubble bursts, hypes, and deflations, but from a technological point of view, the latest wave, almost exactly one year ago set forth by Open AI, but also other, technology companies in terms of large language model in my opinion is an inflection point of the the capability of this technology but it's also an

inflection point of the public awakening including policy circle awakening I want to ask you one last question if you can take out your crystal ball I mean there are people alive now who are over a hundred years old some of them are 110 and their lifetime of span just about all of modern physics, right? Going back to Einstein and relativity and quantum mechanics and black holes. Can I have you take out this crystal ball? Maybe not look so far ahead 100 years from now,

but maybe when I have you back in that seat, 10 or 15 years from now. I hope I get back earlier than that. Okay. Well, tell me why you would be back earlier and tell me what would be happening to bring you back earlier and where would you see things going. Well, I do think AI is a transformative force,

in our society's change, upcoming change, and the continued dialogue and exchange of ideas with the public is very important. I do believe this technology will continue to progress. We have seen the language-based models getting more and more incredible. But we also are going to see multimodal, we're going to see vision and videos. We're going to get into more robotic advancements. You know, all this is part of AI's future. Well, we have run out of time.

I'm so happy to have you as a guest and to talk with you about all of this. Thank you, Erin. You're welcome. Dr. Fay-Fei Lee is the author of the new book, The World's I See, Curiosity, Exploration, and Discovery at the Dawn of AI. She's also the founding director of the Human Centred AI Institute that's at Stanford University, based in Stanford, California. That's it for today. tomorrow, we're debunking the gendered myths of hunter-gatherers. I'm sci-fri producer Shoshana Booksbaum.

See you soon.