Thinking
Nets
review
by Hans Moravec
Brainmakers:
How
scientists are moving beyond computers to create a rival to the human
brain
by David H. Freedman
Simon&Schuster, April 10, 1994
(200? pages)
World War II gestated two kinds of brain-like machine,
analog computers to direct fire
at airplanes and bombs from them, and digital computers to break codes, prepare
tables and simulate atomic explosions. Analog computers acted like small bits
of nervous system, digital computers mirrored mental abstractions. The approaches
spawned independent post-war academic disciplines explicitly aimed at artificial
thought, named by Cybernetics by Norbert Wiener, and Artificial Intelligence
by John McCarthy. By 1960, two former classmates had become the rival
spokesmen for the two approaches.
Frank Rosenblatt at Cornell championed pattern-recognizing electronic nerve
nets called the Perceptrons, while Marvin Minsky at MIT encouraged students to
write thought-simulating programs for digital computers. In 1957 Sputnik, implying
intercontinental nuclear missiles, put fear of technological surprise into
the US defense community and encouraged it to substantially support far out university
research. Thinking machines were a major beneficiary, but fixed funding
totals created serious competition
between research groups. Already a bit stagnant after a promising start,
neural net research screeched to a stop with publication of Minsky's and Seymour
Papert's Perceptrons, which proved fundamental limitations in
two-layer nets, and convinced DOD research managers to divert funding to reasoning
programs--notably at Minsky's lab.
Reasoning programs themselves
stagnated in the 1970s, after promising initial results and inflated promises.
In 1983 John Hopfield showed
how to train three-layer nets, which overcame many Perceptron limitations, and
by the late 1980s neural nets again had funding and a following. Reversing the
situation of twenty years earlier, some young enthusiasts declared the reasoning
program approach dead. Some reporters believed them.
Freedman, a
science reporter, tells a good historical and technical tale. A few errors in
names and events are confined mostly to the introductory chapter, written after
the main text and apparently
insufficiently fact checked. We learn about logic programs and expert systems
as well as Perceptrons and the neural net "connectionism" of the 1980s. The narrow
successes of the many expert systems in business are contrasted against the
few applications and many promises of neural nets. Reasoning programs follow
chains of inference along if-then rules extracted in tedious interrogations of
human experts, while neural nets automatically learn to classify imponderable
input patterns from examples.
In detecting credit card fraud, for example, nets trained from millions of past
transactions beat expert systems distilled from experienced credit analysts.
Echoing the connectionist enthusiasts, Freedman argues that expert systems, and
computers in general, are a dead end for machine intelligence, but that nets
are up to the job. The book presents Doug Lenat's decade-long "Cyc" effort, to
encode common sense in as many as 100 million expert system rules, as the last
forlorn hope for reasoning programs.
I
was not convinced, for reasons evident in my own field of robotics,
as well as many commercial "intelligent" programs for writing and speech
recognition, symbolic mathematics, language translation, game playing, industrial
vision and more, which use pragmatic mixtures of numerical, statistical,
inference and learning methods that are rarely either expert systems or neural
nets. Expert systems and neural nets are nonspecific programming techniques for
encoding decision criteria and
learning input-output relationships, that are often inferior to more problem-specific
encoding, inference and learning algorithms. It is highly improbable
that future fully intelligent machines will be built with either technique alone,
though both may be used in places. Blind biological evolution may be stuck
with solutions once chosen, but intelligence-guided technology is not so limited.
A weak versions of the book's position is justified. Early thinking programs
were so small that it was possible
to hand tailor every parameter and decision point. As computers and problem
sizes have grown, it has been increasingly necessary and worthwhile to entrust
the details to automatic search and learning processes, of all sorts. The
amount of program construction and tuning delegated to the machine is sure to
continue to grow.
Freedman relays the strange claim that nets will
make computers obsolete, suggesting an image of soft, lifelike nets versus rigid,
mechanical computers. In the
1960s, when computers were fabulously expensive, Perceptrons were analog devices,
but in 1980s connectionists used digital computers grown cheap and powerful.
Computers are, after all, universal machines, and can represent any degree
of "softness" in fine resolution numbers. Most commonly, "nets" are simply matrices
of interconnection weights adjusted by a training algorithm. Other many-parameter
learning programs have similar structures--hidden Markov models learn
matrices of transition probabilities,
and high degree curve fitters tune arrays of coefficients. The few nets
implemented as special chips don't alter this observation: other learning algorithms
can likewise be committed to special hardware. It rarely pays however,
because, in a world of rapidly accelerating universal computers and improving
algorithms, speed advantage is temporary, but algorithm inflexibility costs permanently.
Freedman
is a young reporter swept away by young researchers
intoxicated by a hot idea.
No harm done, but things are bound to look differently in daylight, after the party
ends and the hangovers subside.
________________________________________________________
Hans
Moravec is a Principal Research Scientist with
the Robotics Institute of Carnegie Mellon University. He has been developing
spatial perception for mobile robots for two decades, and promises practical results
by the end of the third. He is author of Mind Children: The
Future of Robot and Human Intelligence
(Harvard 1988) and the forthcoming Mind
Age: Transcendence through Robots (Bantam 1995).