Ben,
I just tossed that out without thinking too much about it - just instinctively
knowing it was interesting.
In fact, what is going on is presumably very profound - & "active learning"
doesn't really describe it at all.
Presumably, the situation is something like this:
the monkey is introduced to an "activity" -
he doesn't know wtf the activity is - something to do with pictures, pressing
them, and the def. good part - getting an MM at the end IF you do the activity
right
now the monkey has to work out what the activity really involves largely by
himself - the rules, or most of the rules, of the game - the rules for
connecting pictures/ presses
and that, roughly, is how humans and probably most animals learn most activities
you start playing every activity without knowing what most of the rules are..
[not so much "active" as "discovery learning"]
there is no AGI that is remotely taking that approach, is there? (Peter Voss is
making noises that way - but hasn't offered any public explanation let alone
"proof-of-concept" (his term) - no?)
but I think that's the heart of true AGI - "discovery" learning automatically
presupposes a general capacity to learn any activity
the question is, of course, what does it take to make a discovery learner?
Mike,
The value of this sort of learning is one of the reasons why I'm so excited
about
rolling out AI systems as "virtually embodied agents" in virtual worlds...
So, I agree that traditionally AI systems have not utilized this kind of active
learning,
but I think that they should, and that there is no in-principle reason why they
can't...
Basically, I think one big reason so few AI researchers have taken this path is
a pragmatic one: building robots is a pain, and even hooking AI systems up to
virtual worlds in a flexible
way is a pain (as the Novamente team is very aware these days, as we're involved
with hooking a very limited subset of the Novamente Cognition Engine up to
Second
Life).
Building systems with simple textual interfaces is easier for lone researchers
or small
teams than building systems that hook up to worlds enabling robust perception
and action
(preconditions for active learning). So, the scant funding for AGI-ish
projects has pushed
AGI researchers away from embodied learning, even if they see the conceptual
value of
it...
-- Ben G
On 8/2/07, Mike Tintner <[EMAIL PROTECTED]> wrote:
[Maybe people would like to compare & contrast this "active learning" with
the way AI systems currently learn]
Monkeys Learn in the Same Way as Humans, Psychologists Report
A
rhesus monkey chooses between images on a touch-screen computer monitor.
In a new study monkeys were asked to select five photographs in a
particular order. Credit: UCLA] A rhesus monkey chooses between images
on a touch-screen computer monitor. In a new study, monkeys were asked
to select five photographs in a particular order. Credit: UCLA
Monkeys seem to learn the same way humans do, a new research study
indicates.
"Like humans, monkeys benefit enormously from being actively involved in
learning instead of having information presented to them passively,"
said Nate Kornell, a UCLA postdoctoral scholar in psychology and lead
author of the study, which appears in the August issue of the journal
Psychological Science. "The advantage of active learning appears to be a
fundamental property of memory in humans and nonhumans alike."
In Kornell's study, conducted when he was a psychology graduate student
at Columbia University, two rhesus macaque monkeys learned to place five
photographs in a particular order. The photographs were displayed on a
touch-screen computer monitor similar to those found on ATMs. When the
monkeys pressed a correct photograph, a border appeared around it. When
either monkey pressed all five photographs in the correct order, he
received a food reward. The chance of guessing all five accurately is
less than one percent.
In all, each monkey learned to order at least 18 separate series of
photographs, which included such items as a fish, a human face, a
building, a football field and a flame from a match. They underwent
three days of training before being tested.
In some of the training trials, the monkeys had to figure out the
correct order themselves, while in others, they had the option of
getting help by pushing an icon in the corner of the screen that caused
the border of the correct photograph to flash. They were rewarded with
an M&M candy each time they correctly completed the task without help
and with a less desirable food pellet when they completed the task with
hints from the help icon. After three days, the monkeys were tested
without the benefit of the help icon.
"Both monkeys did much better if they had studied without a hint than if
they had studied with a hint," Kornell said. "The monkeys did much
better on the first three days when they had the help than when they
didn't, but on the test day, it completely reversed. When they studied
with the hint, there is no evidence they learned anything about the
list. They learned the lists when they didn't get the help."
The findings are closely related to findings in humans that recalling
answers from memory enhances long-term learning.
"The findings were somewhat unintuitive, because passively using the
hint appeared to enhance performance during the study phase of the
experiment but had a deleterious effect on long-term learning," Kornell
said.
What are the implications for human learning?
"Many people incorrectly assume the better you do as you're studying,
the more you're learning," said Kornell, who works in the laboratory of
Robert A. Bjork, professor and chair of psychology at UCLA. "If students
don't test themselves when they read a chapter, they can easily think
they know the material when they don't. When you test yourself as you
study, you may feel like you're making it harder on yourself, but on the
test, you will do much better. Robert Bjork calls this 'desirable
difficulty.' If you want to learn something well, when you're reading,
stop and think about what you've read, and test yourself; you learn by
testing yourself. If you make it more difficult for yourself while you
study, you feel like you're doing worse, but you're learning more.
"Active learning is important in humans and — this study
demonstrates — in monkeys as well," he added.
Less effective passive learning includes listening to a presentation and
reading without testing yourself or summarizing what you have learned.
"When you summarize the material in your own words, that's much more
active," Kornell said. "You can't do that if you don't understand it."
Cramming right before a test does not work as well as spacing studying
out over a longer period of time, Kornell added, citing other research
on learning and memory.
Kornell's research, supported by the National Institute of Mental
Health, was conducted with Herbert Terrace, a professor of psychology at
Columbia. The two monkeys, Macduff and Oberon, are housed at the New
York State Psychiatric Institute, where Terrace has a joint appointment.
Neither animal was harmed in the study, and they were fed daily
regardless of how they performed in the trials.
"Many people," Kornell noted, "have had the experience of listening to a
computer instructor open a menu and go through a series of steps. Then
you try to do it, and you don't even know which menu or what the first
step is. If you are passively following along, you won't remember it as
well as if you're forced to do it yourself. Active learning is much
harder, but if you can do it successfully, you will remember it much
better in the long run.
"If you're learning to serve a tennis ball, you won't get much out of an
instructor taking your arm and practicing the swing over and over," he
said. "That's not going to help you nearly as much as if you serve the
ball yourself."
The situation is the same for monkeys, according to Kornell.
"The way the monkeys learn to remember the correct answers is through
active learning, like humans," he said. "They have to generate the
answers themselves from memory. Generating the correct sequence from
memory resulted in more long-term learning than the more passive
training with hints."
Kornell noted that more than a century ago, author William James
remarked on the importance of being actively involved in learning. Since
then, science has proven him correct. Kornell also noted that his
research confirms the teachings of another monkey: Curious George.
Source: UCLA http://www.physorg.com/news105198144.html
<http://www.physorg.com/news105198144.html>
----------------------------------------------------------------------------
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?&;
------------------------------------------------------------------------------
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?&;
------------------------------------------------------------------------------
No virus found in this incoming message.
Checked by AVG Free Edition.
Version: 7.5.476 / Virus Database: 269.11.0/929 - Release Date: 31/07/2007
17:26
-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&id_secret=27945306-82bae8
