1. Flowers
In 1936 the statistician Ronald Fisher published a paper called “The Use of Multiple Measurements in Taxonomic Problems,” in the second issue of volume seven of the Annals of Eugenics. In it was a table: three species of iris, fifty flowers of each, a hundred and fifty flowers altogether. Four measurements for each flower: sepal length and width, petal length and width.
Fisher names the person who measured the flowers: the botanist Edgar Anderson. Anderson collected two of the species from the same patch of ground, fifty of each. The year before, he had published a short piece called “The Irises of the Gaspé Peninsula.” The peninsula is in eastern Canada, jutting into the Gulf of St. Lawrence.
Fifty-one years later, in 1987, a doctoral student called David Aha set up a file-transfer archive on a computer at the University of California, Irvine. He put in datasets that several PhD students there had already collected, along with descriptions, so people outside the university could come and get them. In a 2019 recollection he said there were only a few datasets at first. The archive later became known as the Machine Learning Repository. Anderson's hundred and fifty flowers found their way into it.
A botanist on a peninsula, crouching among wild flowers, measuring one after another. The six hundred numbers became the first exercise that generation after generation of machines did when they learned to classify things.
2. Boxes
Around 1990 the National Institute of Standards and Technology needed a machine to recognize handwriting for a census project. First it needed some handwriting. Forms went out to 3,400 permanent field staff at the Census Bureau; 2,100 came back. On each form were little printed boxes. The person filling it in wrote one character to a box. Afterward a program sliced up the grid, making a small image out of each character.
Another batch of forms went to a high school. Five hundred students in mathematics and science classes filled them out and handed them in on the spot. The proceedings from 1992 make a point of noting the difference between the two groups. The Census Bureau staff returned their forms voluntarily, had a professional reason to make an effort, and tended to write more neatly. The students had been asked to finish them then and there as a classroom exercise.
That year the institute held the first Census Optical Character Recognition Systems Conference. The Census Bureau staff's writing was for the competing machines to practice on. The students' writing was for the test; the answers were not given to the entrants beforehand. The question was how much they got wrong. On digits, the best entrant got 1.56 percent wrong and the worst got twenty percent wrong.
3. Human Performance
There is another experiment in the same proceedings. The institute asked one person to recognize the first ten thousand digits in the students' test set. The person didn't know the answers in advance and went through roughly two characters a second, working one or two hours at a stretch and leaving several hours between sessions. The error rate was 1.57 percent. The best machine's was 1.56.
The conference concluded that recognition of isolated handwritten digits was approaching human performance. Two qualifications came right after that in the same passage. The Census Bureau handwriting didn't come from a broad enough sample, and no human comparison under the same conditions had been made for letters. Human performance here meant one person's performance.
Six years later, in 1998, Yann LeCun and three colleagues published a long paper describing how they remade a dataset of handwritten digits. The institute had originally put the Census Bureau staff's writing in the training set and the students' writing in the test. But the Census Bureau writing, the paper says, was much cleaner and much easier to recognize. It also says why: Census Bureau employees wrote the first set, high school students the second. The test was harder than the practice work because the writers changed, not because the questions did.
So they mixed the two batches and split them anew. They divided the five hundred students into two groups by writer, one for training and one as a pool for the test. They then used Census Bureau writing to bring the training set up to sixty thousand images. The official test had ten thousand images, half from students and half from Census Bureau staff. They called it MNIST, for the modified NIST database. For more than twenty years afterward, it was one of the most commonly used tests in the field.
4. The Turk
In November 2005 Amazon launched a platform. A job on the platform was called a Human Intelligence Task. Whoever offered the job was a requester. Whoever did it was a worker.
In 2006 Felipe Cabrera of Amazon Web Services explained the idea to Wired: some things are hard for computers and easy for people. The platform packaged getting people to do small things as a service that software could call directly. Amazon and news reports at the time used a phrase for this: artificial artificial intelligence.
The platform was called Mechanical Turk. It took the name from an eighteenth-century chess machine that went on tour. A figure in Turkish dress sat at a cabinet and played chess against people. It looked like a machine that could play by itself. In fact the moves were made by a person hidden inside the cabinet. It appeared in Vienna around 1770. For decades people wrote about how the trick worked; Edgar Allan Poe wrote a piece about it in 1836. What Amazon borrowed was this part of the story: from the outside it's a machine; on the inside a person is doing the work.
Someone went to ask who worked on the platform. In February 2010, Panos Ipeirotis at New York University ran a survey. Of the respondents, 46.8 percent were in the United States, 34 percent in India, and the rest in sixty-six other countries. He and two colleagues later conducted a survey that ran for twenty-eight months. They received more than eighty-five thousand responses from about forty thousand different people. One finding was that people were continually coming and going. Any one survey caught only a moment.
In late 2017 a group of researchers published an analysis of anonymous logs recorded by a browser extension between September 2014 and January 2017: about 3.8 million tasks done by 2,676 workers. They counted time spent looking for work, reading instructions, and doing work that was rejected without payment. The median hourly pay came to about two dollars. Only about four percent earned more than the US federal minimum wage at the time, $7.25 an hour.
5. ImageNet
At CVPR in June 2009, six authors including Jia Deng and Fei-Fei Li at Princeton published “ImageNet: A Large-Scale Hierarchical Image Database.” Its framework was already there. Since 1985, a group of psychologists and linguists at Princeton had been building an English lexical database called WordNet. Its basic unit was the synset: words with the same meaning went into one set, and relations such as broader and narrower meanings linked the sets. ImageNet took the noun synsets in WordNet, about eighty thousand of them, and set out to provide five hundred to a thousand screened images of the actual things for each one. By the time the paper appeared, it had completed twelve subtrees, comprising 5,247 synsets and about 3.2 million images.
Where did the pictures come from? The team ran the WordNet terms through image search engines, got back piles of pictures, and gave them to people to look through individually. The question was small: is this thing in the picture or not? The people it went to were on Amazon's platform.
By 2014, ImageNet had images for 21,841 synsets and 14,197,122 labeled pictures. A 2016 Stanford retrospective says that nearly fifty thousand workers from 167 countries took part, over about two and a half years.
Fei-Fei Li later told the story of how the project started many times. Each time it was about making the data bigger first, rather than writing a few more rules for looking at pictures. The 2009 paper was assigned to the poster area. It wasn't given as a talk.
6. The Exam Room
Beginning in 2010, ImageNet held an annual Large Scale Visual Recognition Challenge. There were a thousand categories. The training images and their answers were public. The test images were public, but their answers were not. Beginning in 2011, teams submitted their answers to a scoring server, which checked them against the hidden answer key. The organizers said why the key was hidden: it was to keep people from repeatedly tuning their programs on the test. The server remained open after the competition, with at most two submissions per team each week.
The measure was top-five error. The machine could give five guesses; if any was right, it counted as right. NEC won in 2010 with 28.2 percent wrong. In 2011 Xerox Research Centre Europe came in at around 25.8 percent. In 2012 the three-person team at the University of Toronto reached 15.3 percent. In 2013 a company called Clarifai reached 11.7. Google's team reached 6.66 in 2014; Microsoft Research, 3.57 in 2015. In 2016 a team called Trimps-Soushen from the Third Research Institute of China's Ministry of Public Security reached about 2.99. In 2017 the figure was around 2.25 percent.
On June 2, 2015, the organizers issued a statement. Between late November 2014 and mid-May 2015 a Baidu team had made at least two hundred submissions to the server using at least thirty accounts. It had made more than forty in five days in mid-March. The organizers said the results could not be compared directly with those of other teams. They asked Baidu to stop submitting for twelve months. The team apologized for the mistake. Several days later Baidu announced the results of its investigation: a team lead had instructed several junior engineers to exceed the limit of two submissions a week. The company had terminated the team lead's employment.
There was another score in the room: what a person got wrong. On September 2, 2014, a Stanford graduate student named Andrej Karpathy wrote about an experiment on his blog. He first used five hundred validation images to learn the thousand categories, then identified 1,500 test images himself, at about one a minute. His top-five error rate was 5.1 percent. On the same batch Google's network got about 6.8 percent wrong. He qualified the comparison: it was based on the set of images he had actually worked through. The organizers' later paper describes a second person who practiced on just a hundred images, identified 258, and got twelve percent wrong.
On February 6, 2015, Kaiming He and three colleagues at Microsoft Research published a paper with the words “Surpassing Human-Level Performance” in the title. They had reached 4.94 percent. The human level they compared it with was Karpathy's 5.1 percent. The paper qualifies the comparison with “to our knowledge.”
7. The People on the List
In October 2006 Netflix announced a prize for a recommendation algorithm. Anyone who could improve on its own rating predictions by reducing the error by a tenth would get a million dollars. The public training set held about 100.48 million ratings from 480,189 subscribers for 17,770 films. Each record gave a rating from one to five stars and a date. The subscribers' names had been replaced with numbers.
That year Arvind Narayanan and Vitaly Shmatikov of the University of Texas wrote a preprint that appeared formally two years later. They weren't uncovering names in the file itself: they compared the anonymous records with film ratings and approximate dates people had left on another public site. The point was that even sparse records could be linked back to people if there was other information alongside them.
On September 21, 2009, the million dollars went to a team that had been formed by merging several teams. In December of that year a lawsuit challenged the release and a planned second contest, alleging that the anonymous ratings could still be traced back to people and reveal what films they had watched. In March 2010 Netflix announced it would not hold a second contest. In 2012 Netflix engineers wrote on the company's blog that the new methods late in the competition had brought real improvements in offline tests, but the extra accuracy didn't seem to justify the engineering effort of putting them into use.
ImageNet has a large category called person. The team hadn't drawn up this category from scratch. It inherited the whole tree from WordNet's hierarchy of meanings. In September 2019 Kate Crawford and Trevor Paglen launched a project called Excavating AI. There were more than 2,800 categories under person. An accompanying site, ImageNet Roulette, let people submit a picture of themselves and returned a label that the dataset placed under person.
In January 2020 the ImageNet team itself published a paper proposing to remove 1,593 labels for subcategories under person that were derogatory, offensive, or involved sensitive identities, along with the associated 600,040 images. In June that year MIT researchers announced that they were formally withdrawing a separate, earlier dataset, 80 Million Tiny Images, and asked people who had used it to delete their copies. Around that time Vinay Prabhu and Abeba Birhane published a paper taking stock of derogatory categories and pictures of people used without consent in large image datasets of this kind.
There was another test about how machines see people. In February 2018 Joy Buolamwini and Timnit Gebru published “Gender Shades.” They made the test set themselves. About 1,270 public photographs showed members of parliament from three African and three European countries, grouped by gender and by the six-part scale for skin type used in dermatology. They tested publicly available gender-classification services from three companies. For women with darker skin the error rates were about 34.7, 34.5, and 20.8 percent. For men with lighter skin all three were below one percent. Two of the companies subsequently released improvements and said their error rates had fallen for people with darker skin.
8. One Person
Human level appears twice in this essay. In 1992 it was one person, ten thousand digits, two a second, one or two hours at a stretch. In 2014 it was also one person: 1,500 pictures, one a minute, after getting familiar with the categories on five hundred practice images. Both people took the work seriously, and both accounts give the way they did it. What traveled were two sentences: the machine came close to people; the machine passed them.
What was cut between one person and people in general was Karpathy's qualification on his blog, the words Kaiming He and his colleagues added, and the two qualifications following the 1992 conclusion. With those gone, what one person scored on one test became a line for the human race.
People stand behind the test papers too. The botanist measuring flowers on the Gaspé Peninsula. The 2,100 Census Bureau staff who returned a form and the five hundred high school students who filled in boxes in class. The person who recognized two characters a second. The nearly fifty thousand workers on the platform looking at pictures and saying yes or no. The 480,000 people whose movie ratings were filed under numbers. The 600,000 or so images included under person that the team later proposed taking down. The more than a thousand members of parliament in six countries.
One test gave way to the next and the error rate went down year by year. By the end, what the teams competed over was how many more answers they could get to match the key kept out of sight. To match a few more, one team submitted more than forty times in five days. A score is a ruler for measuring a machine. When the ruler itself becomes the thing to reach for, the organizers have to hide the answers and allow no more than two attempts a week.
This series calls the part that remains outside, however much is measured, the remainder. It isn't hard to find here. Every character, flower, picture, and rating in these tests came from a hand, a pair of eyes, one person's choice. The machine's score is on the leaderboard. Some of the people left a name. Most left a number, a box, or an hourly wage.
The machines are still being built. Each generation says again where the boundary lies.