et al.
et al.
et al.
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
BOOKS ET AL.<br />
760<br />
senting hundreds of millions of pages such as<br />
the full text of Wikipedia. Y<strong>et</strong> except for these<br />
restricted speci<strong>al</strong> domains, AI remains a long<br />
way from its human variant.<br />
Kurzweil introduces the reader to one of<br />
AI’s signature mathematic<strong>al</strong> techniques that<br />
he himself used to build speech recognition<br />
systems: hierarchic<strong>al</strong>, hidden Markov models<br />
(HHMMs). These are trainable <strong>al</strong>gorithms<br />
well matched to de<strong>al</strong> with the many levels of<br />
human language, from phonemes to words<br />
to sentences and beyond. For certain types<br />
of simple problems, HHMMs can be trained<br />
fairly automatic<strong>al</strong>ly and are easy to implement<br />
in “units” that are connected by ascending<br />
and descending links.<br />
It is here that Kurzweil leaves the solid<br />
ground of his expertise to wade into the muddier<br />
waters of biology and psychology. For<br />
he now categoric<strong>al</strong>ly asserts that HHMMs<br />
provide a powerful and univers<strong>al</strong> model of<br />
neocortic<strong>al</strong> computation, with 300 million<br />
pattern recognizers distributed across the<br />
cortic<strong>al</strong> she<strong>et</strong>.<br />
Kurzweil introspects into his own mind’s<br />
capabilities and the nature of perception and<br />
memory, without bothering to refer to the<br />
massive literature on these topics. One of the<br />
most basic lessons of psychology is that we<br />
have little idea of what goes on in our minds,<br />
as evolution has not given us access to most<br />
parts of the brain (explaining why so much<br />
philosophy of mind has been barren when <strong>al</strong>l<br />
it could rely on was introspecting philoso-<br />
BROWSINGS<br />
phers). He then moves on to the brain, describing<br />
its anatomy in terms of the “new” brain,<br />
higher-order regions of the neocortex, and the<br />
“old” brain (everything else). According to<br />
Kurzweil, the new brain is clever, learns fl exibly,<br />
and controls the primitive impulses of<br />
the old brain relating to food, sex, and aggression.<br />
His understanding of neuroanatomy is<br />
about as sophisticated as U.S. Secr<strong>et</strong>ary of<br />
Defense Don<strong>al</strong>d Rumsfeld’s understanding<br />
of internation<strong>al</strong> politics when he articulated<br />
his belief of a division of Europe into an Old<br />
and a New one during the run-up to the second<br />
Gulf War in 2003.<br />
Kurzweil’s knowledge of neuroscience is<br />
simply inadequate to the task at hand. From the<br />
tens of thousands of studies published annu<strong>al</strong>ly,<br />
he selectively cites a handful of papers<br />
that buttress his points, without giving any<br />
context. He mistakes the striatum for cortex<br />
and apic<strong>al</strong> dendrites for axons, belies the cognitive<br />
contributions of the bas<strong>al</strong> ganglia, and<br />
denies higher ment<strong>al</strong> abilities to insects, ceph<strong>al</strong>opods,<br />
and birds that don’t have a neocortex.<br />
Y<strong>et</strong> he has the unerring belief of the proph<strong>et</strong><br />
(or the fool): “I maintain that the model I have<br />
presented is the only possible model that satisfi<br />
es <strong>al</strong>l of the constraints that the research and<br />
our thought experiments have established.”<br />
Scientists trying to simulate the mindbrain<br />
f<strong>al</strong>l <strong>al</strong>ong a continuum, ranging from<br />
extreme biologic<strong>al</strong> chauvinism (the need to<br />
consider every ionic channel, action potenti<strong>al</strong>,<br />
and neuron to fully do justice to the<br />
Mathematic<strong>al</strong> Excursions to the World’s Great Buildings. Alexander J. Hahn. Princ<strong>et</strong>on University<br />
Press, Princ<strong>et</strong>on, NJ. 2012. 344 pp. $49.50, £34.95. ISBN 9780691145204.<br />
While touring examples of western architecture<br />
from the great pyramids of Giza<br />
to Frank Gehry’s Bilbao Guggenheim<br />
Museum, mathematician Hahn intertwines<br />
two historic<strong>al</strong> narratives: The architectur<strong>al</strong><br />
focuses on aspects of appearance (shapes,<br />
symm<strong>et</strong>ry, and proportion) and mechanics<br />
(loads, compressions, tensions, and<br />
thrusts). The mathematic<strong>al</strong> progresses from<br />
Euclidean geom<strong>et</strong>ry and trigonom<strong>et</strong>ry to<br />
basic c<strong>al</strong>culus. His discussions reve<strong>al</strong> how<br />
mathematics provides insights into the<br />
design and construction of the buildings<br />
and how the buildings incarnate the math.<br />
[Carefully composed using circles and squares, Leon Battista Alberti’s 15th-century facade (above)<br />
for the mediev<strong>al</strong> church Santa Maria Novella in Florence was a source of inspiration to a number of<br />
Renaissance architects.] As Hahn notes, his two strands are linked topic<strong>al</strong>ly rather than chronologic<strong>al</strong>ly,<br />
because the actu<strong>al</strong> builders gener<strong>al</strong>ly lacked the mathematics that clarifi es understanding of their<br />
constructions. Readers wishing to shore up their own understanding of that mathematics can work<br />
through the chapter-end problem s<strong>et</strong>s and discussions.<br />
15 FEBRUARY 2013 VOL 339 SCIENCE www.sciencemag.org<br />
Published by AAAS<br />
baroque complexity of the brain’s circuits) to<br />
the austerity of a purely <strong>al</strong>gorithmic approach<br />
of replicating the mind in software (the mind<br />
is not w<strong>et</strong>, after <strong>al</strong>l). Mathematicians and<br />
engineers natur<strong>al</strong>ly belong to the later camp,<br />
convinced that one <strong>al</strong>gorithm rules them<br />
<strong>al</strong>l. Previous favorites include logic<strong>al</strong> c<strong>al</strong>culus,<br />
neur<strong>al</strong> n<strong>et</strong>works, cellular automata, selfreferenti<strong>al</strong><br />
programs, and, yes, hidden<br />
Markov models [as proposed a few years earlier<br />
by fellow entrepreneur and inventor of the<br />
P<strong>al</strong>mPilot Jeff Hawkins ( 1)].<br />
Kurzweil correctly points out that the<br />
pace at which biologists accumulate data<br />
has increased dramatic<strong>al</strong>ly over the years<br />
(<strong>al</strong>though I only wish that his claim that “the<br />
spati<strong>al</strong> resolution of brain scanning … [is]<br />
doubling every year” were true). From this<br />
he infers that a compl<strong>et</strong>e understanding of the<br />
brain and the mind can’t be far away.<br />
Paradoxic<strong>al</strong>ly, the endless data fields<br />
make it ever more diffi cult to distinguish the<br />
sign<strong>al</strong> from the noise. Indeed, the torrent of<br />
data beg<strong>et</strong>s the illusion of progress. While<br />
data about the brain accumulate exponenti<strong>al</strong>ly,<br />
our understanding increases sublinearly.<br />
Basic questions about cortic<strong>al</strong> circuitry<br />
posed by future Nobel laureates David Hubel<br />
and Torsten Wiesel in a celebrated publication<br />
in 1962 ( 2) remain unanswered 50 years<br />
later. Function<strong>al</strong> human brain imaging has<br />
y<strong>et</strong> to affect standard medic<strong>al</strong> practice (the<br />
upcoming fi fth edition of the Diagnostic and<br />
Statistic<strong>al</strong> Manu<strong>al</strong> of Ment<strong>al</strong> Disorders does<br />
not even mention any function<strong>al</strong> magn<strong>et</strong>ic<br />
resonance imaging diagnostic criteria). And<br />
even the lowly roundworm Caenorhabditis<br />
elegans, a creature no bigger than the l<strong>et</strong>ter<br />
l and with exactly 302 nerve cells, is for now<br />
beyond the ability of computation<strong>al</strong> neuroscience<br />
to comprehend. Kurzweil’s claim that<br />
we will soon fi gure out how the 100 billion<br />
neurons of the human brain function on the<br />
basis of designed HHMMs is compl<strong>et</strong>e bosh.<br />
One thing is certain. Biology knows nothing<br />
of simplicity. Brains are not assembled<br />
out of billons of identic<strong>al</strong> LEGO blocks but<br />
out of hundreds of distinct nerve cell types.<br />
Each cell type has its own idiosyncratic morphology,<br />
sign<strong>al</strong>ing, and active genes. And<br />
they are interconnected with elaborate wiring<br />
rules that we only discern darkly. To paraphrase<br />
Winston Churchill, neuroscience is<br />
(perhaps) at the end of the beginning of the<br />
quest to understand our brain and mind.<br />
References<br />
1. J. Hawkins, S Blakeslee, On Intelligence (Times Books,<br />
New York, 2005).<br />
2. D. H. Hubel, T. N. Wiesel, J. Physiol. 160, 106 (1962).<br />
10.1126/science.1233813<br />
CREDIT: JEBULON/WIKIMEDIA COMMONS<br />
on February 14, 2013<br />
www.sciencemag.org<br />
Downloaded from