Computer learns like a human and (sort of) beat the Turing Test
Computer learns like a human and (sort of) beat the
Turing Test
BY LANCE ULANOFF 1 DAY AGO
There is a fundamental difference between the way
computers learn and the way humans learn. Humans can see one example and intuit
what that object or symbol might be used for and quickly identify similar things.
A computer can only arrive at the same conclusions after being fed thousands
and thousands of examples. This usually referred to as "machine
learning."
That may, however, be about to change.
Scientists at New York University have figured out a way
to not only mimic how humans make those mental leaps, but to have computers
recreate simple symbols and drawings in such a way that they're almost
indistinguishable from these created by humans.
In a paper published this week in Science researchers
describe how they built a “Bayesian Program Learning (BPL)” algorithm, which
turns concepts into simple computer programs and allows computers to learn a
large class of visual concepts from a single example.
In a digital alphabet, the letter “A” would be
represented by code. However, instead of a programmer writing the code, the
computer generates the code to represent the letter and then it also produces
variations based on that first letter.
Researchers said the model uses knowledge from previous
concepts to learn. For example, if the computer knows the Latin alphabet, that
can help it learn the similar Greek alphabet.
Brenden Lake is a Moore-Sloan Data Science Fellow at New
York University and the paper's lead author. He says the breakthrough came when
researchers noticed that, “If you ask a handful of people to draw a novel
character, there is remarkable consistency in the way people draw.... They do
not see characters as just static visual objects. Instead people see richer
structure... that describes how to efficiently produce new examples of the
concept.
"We aimed to develop an algorithm with the same
capability and then compare it with people.”
During a presentation on their work, the scientists said
they've not only built a machine-learning program, “but what the program learns
— its concepts — are also programs. We think that is true for humans too: your
concepts are programs, or parts of programs,” said Joshua Tenenbaum, of the
Department of Brain and Cognitive Sciences, and Center for Brains, Minds, and
Machines, at MIT.
More incredibly, when the computer was asked to create
fresh examples based on the original concept, and those images were compared to
examples created by humans, humans often couldn’t distinguish if a person or a
computer had created the example.
In other words, the computational model passed a rough
form of the Turing Test. Legendary 20th century mathematician Alan Turning (he
broke the Enigma Code) posited that in the 21 century a computer would have a
70% chance of fooling a human into thinking they were communicating with
another human.
According to the study, “this approach can perform
one-shot learning in classification tasks at human-level accuracy and fool most
judges in visual Turing tests of its more creative abilities. For each visual
Turing test, fewer than 25% of judges [the paper notes that there were 35]
performed significantly better than chance.”
"Our results show that by reverse-engineering how
people think about a problem, we can develop better algorithms," Lake said
in a release.
This machine-learning shortcut could have wide-reaching
implications. It could shorten the time it takes for computers to learn new
languages, recognize images and help systems generate new, usable designs based
on existing designs, without human input.
The research could also have significant impact on future
artificial intelligence innovation, including robotics. A robot that can make
logical leaps about things might someday be more adept at human-like
decision-making. Which is either a thrilling or terrifying thought.
Not everyone agrees that this is a breakthrough or even
that the system beat the Turing test.
Allen Institute for Artificial Intelligence CEO Oren
Etzioni told Mashable that "They didn’t beat the Turing test any more than
a calculator does by out-multiplying a human," and the work is best
classified as a "scientific contribution."
"While the authors pose a fascinating research
question, many researchers have used related methods to achieve strong results.
Still, the paper is an invaluable reminder that we need methods that can
generalize from small numbers of examples both to model human abilities and to
move AI forward," said Etzioni.
Even if you do buy into the idea that this is a
significant advancement in the field of AI, applications for the work are
years, if not decades away; even by the researchers' own measure, the program
still doesn’t see the same level of structural detail as humans. “It lacks
explicit knowledge of parallel lines, symmetry, optional elements such as cross
bars in '7's, and connections between the ends of strokes and other strokes,”
wrote the scientists.
Comments
Post a Comment