My latest Mind and Matter column is on Ray Kurzweil’s
new book:
When an IBM computer program called Deep Blue defeated
Garry Kasparov at chess in 1997, wise folk opined that since chess
was just a game of logic, this was neither significant nor
surprising. Mastering the subtleties of human language, including
similes, puns and humor, would remain far beyond the reach of a
computer.
Last year another IBM program, Watson, triumphed at just these
challenges by winning “Jeopardy!” (Sample achievement: Watson
worked out that a long, tiresome speech delivered by a frothy pie
topping was a “meringue harangue.”) So is it time to take seriously
the prospect of artificial intelligence emulating human
abilities?
Yes, argues the inventor and futurist Ray Kurzweil in his new
book “How to Create a Mind.” Mr. Kurzweil reckons
that a full understanding and simulation of the human brain is a
lot closer than most people think. Since he has a more impressive
track record of predicting technological progress than most, he
deserves to be heard.
It’s become fashionable to think of the brain as so intricate as
to be almost beyond even theoretical comprehension. For
example, Paul Allen, the Microsoft co-founder,
criticized both IBM’s Watson and Mr. Kurzweil in a recent article,
claiming that the former’s knowledge was brittle and
domain-specific, while the latter failed to understand that every
structure in the brain “has been precisely shaped by…evolution to
do a particular thing.” Mr. Allen posits a “complexity brake” that
would necessarily limit the understanding and replication of the
brain.
Mr. Kurzweil’s reply in his book is persuasive. For a start, the
brain is built from a relatively small and simple body of
information-the 25 million bytes of the genome. The complexity
comes from ordered growth and elaboration. Second, the brain
contains massive redundancy, with certain kinds of basic
pattern-recognizing circuits repeated maybe 300 million times in
different brain regions. Third, as Van Wedeen of Harvard Medical
School and colleagues found in a recent study, much of the brain
has a horizontal grid of fibers running at right angles, connecting
vertically: a bit like the streets and elevators of Manhattan.
Moreover, the design of artificial intelligence systems has been
converging with the way brains developed. Using evolutionary
algorithms (a fancy form of trial and error), Mr. Kurzweil himself
developed some of the successful speech-recognition software that
we all take for granted.
Mr. Kurzweil agrees with another innovator turned
neuroscientist, Jeff Hawkins (the PalmPilot’s inventor), in
believing that the human brain is basically a set of prediction
machines that work by forecasting how a pattern of perceptions will
develop. As we put together the pieces of, say, a visual image,
information is flowing up (by the neural grid’s elevators) from
basic pattern recognizers to higher and more abstract integrations,
but also back down from the higher levels predicting what patterns
will be found in missing parts of the image or as an image changes.
Failed predictions-“surprises”-may be passed (via the neural grid’s
streets) to higher levels in the neural hierarchy for conscious
resolution.
If this picture is broadly right, then replicating a brain isn’t
impossible. Engineers trying to reduce circuitry features to five
microns once scoffed at ever reaching one micron, while today
they’re at 0.022 microns. So, says Mr. Kurzweil, the
neuro-pessimists will be wrong because “the project to
reverse-engineer the human brain is making similar progress.”
Moreover, says Mr. Kurzweil, the brain is an essentially linear
organ, sequentially processing information, which may be why we
find it so hard to comprehend the nonlinear trends that so dominate
the progress of technology. With both hardware and software
changing exponentially, it would be foolish, not wise, to bet
against the emulation of the human brain in silicon within a couple
of decades.