Return of the human computers
From
The Economist, December 3 2011
It was late summer 1937, and the recovery from the Depression had stalled.
American government officials had stimulus money to spend but, with winter
looming, there were few construction projects to fund. So the officials created
office posts instead. One project was assigned to a floor of a dusty old New
York industrial building, not far from Times Square. It would eventually house
300 computers—humans, not machines.
The computers crunched through the calculations necessary to create
mathematical tables, then an indispensable reference tool for many scientists.
The calculations were complex and the computers, drawn largely from the ranks
of New York’s poor, possessed only basic numeracy. So the mathematicians in
charge of the project worked out how to break each calculation down into simple
operations, the outcomes of which could be combined to give a final result.
It was a technique that had been employed for decades across America and
Europe. The field of human computing even had its own journal and trade-union
representation. Computing offices calculated ballistics trajectories, processed
census statistics and charted the course of comets. They would continue to do
so until the 1960s, when electronic computers became cheap enough to consign
the profession to history.
Until recently, that is. Over the past few years, human computing has been
reborn. The new generation of human computers carry out different tasks, but
they mirror their predecessors in many other ways. They are being drafted in to
perform tasks that computers cannot. They are employed in large numbers and are
organised into streamlined workflows. And, as was the case in the age before
electronic computers, their output is combined to generate results that could
not easily be produced in any other way.
In one proof-of-principle experiment, published earlier this year, human
computers were used to create encyclopedia entries. Like performing
mathematical calculations, this is a skilled job, but one that can be broken
down into simpler parts, such as initial research, writing and editing. Aniket
Kittur and colleagues at Carnegie Mellon University in Pittsburgh, Pennsylvania
created software, known as CrowdForge, that manages the process. It hands out
tasks to online workers, which it contacts via Mechanical Turk, an outsourcing
website run by Amazon. The workers send their work back to CrowdForge, which
combines their output to produce surprisingly readable results.
Several American start-ups are operating similar workflows. CastingWords
breaks audio files down into five-minute segments and farms each out to a
transcriber. Each transcription is automatically bounced back to other workers
for checking and, once deemed good enough, an (electronic) computer combines
the segments and returns the finished product to the customer. At CloudCrowd a
similar system is used to co-ordinate teams of human translators. Others are
combining human and artificial intelligences. An app called oMoby, produced by
IQ Engines, can identify objects in images snapped by iPhone users. First it
applies object-recognition software, which may not be able to cope if the
lighting is poor or the image was captured from an unusual angle. When that
happens, the image is sent to a human analyst. Either way, the user gets an
answer in half a minute or so.
Much more is to come. In old-fashioned computing offices, workflows were
co-ordinated by senior staff, often mathematicians, who had worked out how to
deconstruct the complex calculations the computers were tackling. Now silicon
foremen such as CrowdForge oversee human computers. These algorithms, which
co-ordinate workers by plugging into Mechanical Turk and other online piecework
platforms, are relatively new and are likely to get considerably more
sophisticated. Researchers are, for example, creating software to make it
easier to assign tasks to workers—or, to put it another way, to program humans.
Eric Horvitz, a researcher at Microsoft’s research labs in Redmond,
Washington, has considered how such software could be put to use. He imagines a
future in which algorithms co-ordinate an army of human workers, physical
sensors and conventional computers. In the event of a child going missing, for
example, an algorithm might assign some volunteers to search duties and ask
others to examine CCTV footage for sightings. The system would also trawl local
news reports for similar cases. These elements would be combined to create a
cyborg detective.
This sounds terribly futuristic, and rather different to the pen-and-paper
human computation of the 19th century. But David Alan Grier, a historian of
computing at George Washington University in Washington, DC, thinks that the
architects of the new systems could learn a lot by studying the old ones. He
points out that Charles Babbage, the designer of an early mechanical computer,
gave much thought to reducing the errors that human computers made. Babbage
realised that duplicating tasks and comparing the results was not enough,
because different workers tended to make the same mistakes. A better solution
was to find different ways to perform the same calculation. If two methods
produce the same answer, the result is much less likely to be flawed, Babbage
reasoned.
There are many more such useful tips in the historical record, says Dr
Grier. Human-computing pioneers also wrote a lot about how best to break a
complex calculation into sub-tasks that are completely independent of each
other, for example. “There are all sorts of hints in the old literature about
what’s useful,” he says. He is often invited to human-computing conferences at
which he likes to chide researchers for overlooking such lessons from this
forgotten but intriguing early chapter of computer history.