Using AI to Increase Your Intelligence & Enrich Humanity | Dr. Fei-Fei Li

The episode opens with Fei-Fei Li's view of generational anxiety about technology. She argues that the thing humanity "never learns" is the older generation lamenting that the next one knows nothing, is rude, or is forgetting the past. Looking at the arc of history, she says humanity has, "by and large," advanced for the better. She explicitly does not deny the atrocities or the setbacks. Children, in her account, are curious: they are "massively entertained" by the technology, but they are also starting to use it.

What I worry about are teachers and some parents, because I think our society today, and especially Silicon Valley, are not doing them a service. We're forgetting about them.

Andrew Huberman introduces Li as a Stanford computer scientist and professor, a pioneer of artificial intelligence and computer vision, and director of the Stanford Institute for Human-Centered Artificial Intelligence. He notes that she has been called the godmother of AI, both for helping bring in AI technologies and for insisting that ethics and benevolent uses stay central to AI and robotics. He sets the conversation up as an examination, from a neuroscience perspective, of what intelligence is, how human brains learn, and how AI, being built on internet content, both resembles human learning and falls short of it. He adds that Li acknowledges the many valid concerns people have about AI.

Huberman starts with vision, observing that for most people vision and light sound unrelated to AI, though he holds that this is where it all starts. Li answers that vision is a cornerstone of intelligence along two parallel lines: biological evolution, and the history of computer vision within AI.

On the evolutionary side, Li says that 540 million years ago animals "saw the first light." These were simple sea animals such as trilobites and their relatives. Before then there was very little sensing. Touch was starting to emerge around the same time, but by her account there was no hearing, no smell, and no nervous system at all. Her argument is that photoreceptive cells created an evolutionary force. Sensing the outside world changes an animal's relationship to it: an animal that can see can find food, seek mates, and become someone else's food. She cites fossil studies showing that the Cambrian explosion of animal speciation came 10 million years after the first light.

To put it simply, if you can see food, it changes your life, from a evolution point of view

Li also argues that vision remains central to advanced intelligence. She cites an estimate that half of the cortical activity in the human brain is involved in visual function, and notes that children are visual before they are verbal in development.

Li's first contribution concerns algorithms. Computer scientists started experimenting with neural networks in the early 1950s. In the same period, neuroscientists such as Hubel and Wiesel were recording from visual cells in the mammalian brain. They found a hierarchy of nerve cells that pass information upward, from collecting light at the retina to recognizing that a shape is present. Li says this architecture was part of the inspiration for neural network algorithms. She also marks the limits of the comparison: today's networks run on hundreds of billions or even trillions of parameters, with a complexity that "departs from" what has been recorded in the mammalian visual pathway, even though both lines of work began a little over half a century ago.

The second contribution is data, which is where Li's own work comes in. Around the turn of the century, she says, machine learning was in an exploratory phase. Labs were trying neural networks alongside other methods, such as Bayesian methods and support vector machines. In 2006, as a first-year faculty member at Princeton, she and her students noticed how little data these algorithms were given to learn from. She turned to the cognitive neuroscience literature on vision and found that by age six, humans can learn tens of thousands of object categories. They are also exposed to a huge amount of visual input, since babies can see most of the time from the moment they are born.

So we conjectured that the lack of data was a huge part of the reason there's a lack of progress in AI.

Li says this put her group at odds with researchers focused only on algorithms. The result was ImageNet, which she describes as the first internet-scale dataset for AI. She says it held 50 million images and was meant to drive machines to recognize everyday objects such as microphones, cups and chairs. By her account, ImageNet came together with maturing neural network algorithms and GPU computing, and by 2012 that convergence of three elements "became the defining moment of what modern AI is."

Huberman recalls a debate around 2012 at the vision course held every other summer at Cold Spring Harbor: could a computer learn to recognize specific faces as well as humans? He suggests most people would now say computers are much better at it, even though some "super recognizer" people are exceptional at it, and Li agrees. He asks how the technology went from confusing a person with a lookalike to being precise.

Li goes through the convergence again in more detail. GPU computing parallelized computation, which allowed more operations to pass through the algorithms. Neural networks, which began in the 1950s as stacks of very simple units, each taking an input and applying a simple function, had matured by around 2010. Finally, she describes a "reckoning moment" in which researchers saw that big data, made more available by the internet, had to be part of the equation for learning patterns. She stresses that these artificial units are much simpler than real neurons.

She then describes the ImageNet challenge her lab started in 2010, when GPUs were not yet mature. It was a public, multi-year competition in object recognition. The test data covered 1,000 object categories and more than a million images. An algorithm was shown a picture, had to name the main object, and got a point for a correct answer. A few years later, a graduate student at Stanford was used to benchmark human performance. Li first gives the result as roughly 4%, then corrects herself: 4% was the human error rate. Random guessing would be right one time in 1,000.

Machines did worse than humans in the early years. Li calls 2012 the turning point, when neural networks, the ImageNet data and GPUs came together. That year's error rate was "cut down to the tens," which was still short of human performance. But the drop from earlier algorithms was so large that researchers recognized an inflection point. In her account, algorithms beat humans at naming the 1,000 object categories by 2016. She gives the gap as "another three years" while also citing both 2012 and 2016, so the exact interval is unclear in the transcript.

When Huberman asks where the human 4% error came from, Li says time pressure was not the main issue, though she doubts a graduate student would want to do the task forever. The main limit is memory: holding the patterns of 1,000 object classes is hard even when some classes are familiar. Fine distinctions, such as between closely related dog species, add to the confusion.

Huberman asks whether the same approach was tried with hearing. He notes that humans are excellent at recognizing speech inflection and emotional tone, but that as a non-musician he would struggle to tell apart even 15 different sound frequencies. Li says "the floodgate got open": speech recognition, sound recognition, natural language processing and vision all got a boost. She mentions Stanford colleagues who use machine learning to study whale songs, and says speech recognition did well early in the AI revolution.