Freakonomics Podcast People I (Mostly) Admire, with Steve Levitt, Episode 118
Guest: Fei-Fei Li, professor of computer science at Stanford University specializing in computer vision and a co-director of Stanford’s Human Centered A.I project.
Computer Scientist Fei-Fei Li had a wild idea: download one billion images from the internet and teach a computer to recognize them. She ended up advancing the field of artificial intelligence — and she hopes that will turn out to be a good thing for humanity.
Li offers an overview of her speciality, Computer Vision: since humans use vision as a primary sense to navigate the world, computer vision aims at making machines similarly able to visually recognize objects. She recalls that before 2007, researchers were working with limited data sets of up to 20 different objects (cows, airplanes, beer bottles, etc.) with only a few machine-learning models, and they focused on improving the code rather than data sets. Reading psychology papers that described babies as capable of seeing around 30,000 objects in comparison, Li decided to build ImageNet in 2007, which aimed to solve the problem of object recognition by utilizing vast training datasets.
In 2010 after three years of development, the dataset included 15 million images downloaded from the internet, split into 22,000 visual categories (close to the initial goal of 30,000) – a number that has grown into the billions since. A challenge emerged: how to filter and label the data. Li mentions that if one types the word “German Shepherd,” there is a need to filter between images of actual German Shepherds, German Shepherd posters, German Shepherd cartoons, and random pictures that do not contain German Shepherds at all. She hired undergraduate students to label the data and tens of thousands of online workers from around the globe through Amazon Mechanical Turk. However, she faced reluctance and lack of support from peers and mentors because others either viewed the project as over-ambitious or they disagreed with the methodology, believing it to be more effective to focus on recognition of a single object before attempting multiple object-recognitions.
In 2010, Li launched the ImageNet contest, a competition inviting researchers to test their visual recognition algorithms on a subset of ImageNet’s dataset. She recalls how in the first year, the error rate was 30%. The use of GPUs and convolutional neural networks dropped the error rate to 15% in a matter of months. Li mentions her appreciation for neural networks’ scalability, but is unsure how these models work and why they are so effective.
Li recalls her life journey as a child of immigrants who came to New Jersey in the 90s when she was 15 years old. She went to public high school in ESL classes while working in the neighborhood’s Chinese restaurant kitchen. Later as a first-year at Princeton with a full scholarship, she opened a family dry cleaning business. Due to her parents’ health issues and lack of healthcare insurance, she hesitated to take lucrative jobs in the finance sector, though she was eventually offered a professorship at the University of Illinois Urbana-Champaign. Having chosen to work in A.I. out of curiosity, Li realized during a 2017 sabbatical from her Google Cloud position as a chief scientist of AI that AI could lead to the next industrial revolution and decided to ensure the technology would be here to benefit humanity, not to destroy it.
Recognizing that AI is a double-edged sword that can be appropriated by bad actors or yield unintended consequences even from well-intentioned ones, she joined the Stanford Human-Centered A.I. Institute to facilitate dialogue with policymakers and the public with the hopes of collective oversight in these technologies’ development. She advocates a multi-dimensional regulation of A.I. through law, social norms, and partnerships with cyber-security companies. She envisions the regulation of A.I. similarly to how nuclear power has been regulated by nation states with treaties. She worries mostly about disinformation but finds hope in the efforts by governments, academics, and industry leaders to explore tools such as authentication, watermarking, and social engineering methods. However, the podcast host concludes that Fei-Fei Li hasn’t convinced him that AI can be compared to nuclear bombs. He thinks the regulation of A.I. will require different management strategies, as the barrier to entry in the field of AI is much lower than in the nuclear domain.
