A team of biomedical engineers has developed a
computer model that makes use of more or less
predictable "guesstimates" of human muscle movements to
explain how the brain draws on both what it
recently learned and what it has known for some time to
anticipate what it needs to develop new
motor skills.
The engineers, from Johns Hopkins, MIT and
Northwestern, exploited the fact that all people
show similar "probable" learning patterns and use them to
develop and fine-tune new movements,
whether babies trying to walk or stroke patients
reconnecting brain-body muscle links.
In their report in the June issue of Nature
Neuroscience, the team says that their new tool
could make it possible to predict the best ways to teach
new movements and help design physical
therapy regimens for the disabled or impaired.
Reza Shadmehr, professor of biomedical engineering
at Johns Hopkins, who with his colleagues
built the new model, says that the artificial brain in the
computer, like its natural counterpart, is
guided in part by a special kind of statistical
"probability" theory called Bayesian math.
Unlike conventional statistical analysis, a Bayesian
probability is a subjective "opinion" that
measures a "learner's" individual degree of belief in a
particular outcome when that outcome is
uncertain. The idea as applied to the workings of a brain
is that each brain uses what it already knows
to "predict" or "believe" that something new will happen,
then uses that information to help make it so.
"We used the idea that prior experience and belief
affect the probability of future outcomes,
such as taking an alternate route to work on Friday because
you've experienced heavy traffic Tuesday,
Wednesday and Thursday and believe strongly that Friday
will be just as bad," Shadmehr said. E-mail
spam filters operate on a similar principle; they predict
which key words are "probably" attached to
mail you don't want and "learn" as they go to fine-tune
what they exclude from your in-box.
The computer model, Shadmehr said, almost precisely
duplicates the results of experiments
that tested the ability of monkeys to visually track rapid
flashes of light. Experiments using such
rapid eye movements, or saccades, are a staple in studying
how the brain controls movement.
Initially, the animal learner made large errors but
also stored the information about its
mistakes in a memory bank so it could adapt and make more
accurate predictions the next time around.
Every time the learner repeated the task, it would sift
through the prior knowledge in its memory
banks and make a prediction on how to move, which in turn
would also be memorized. While short-term
memory was periodically purged, repeated errors were
transferred to a long-term memory bank.
The computer learner was tasked with "looking" at a
spot of light, then all the lights were
turned off. The spot of light was turned on again, and the
computer learner was again asked to look at
that same spot. The learner's speed and pattern in adapting
its movements matched the experimental
results of the monkeys almost perfectly. "We found that
this Bayesian model can explain almost all of
the phenomena we observe in regard to learning motor
movements," Shadmehr said.
Beyond possible use in helping stroke patients, the
new tool might also be applied to better
understand how we learn language, develop ideas and make
memories. "How we learn to think operates
under many of the same principles as how we learn to move,"
Shadmehr said.
The research was funded by the Howard Hughes Medical
Institute and the National Institutes
of Health.
Authors on the paper are Konrad Kording, of
Northwestern University; Joshua Tenenbaum, of
MIT; and Shadmehr.