COULD THERE BE REAL-TIME, INSTANTANEOUS LEARNING IN THE BRAIN?

[Asim Roy]

I am posting this memo to various newsgroups. So my apologies if
you get multiple copies.
-----------------------------------------------------------

This is a summary of the responses (comments/questions) I 
have received so far. My sincere apologies for the long
delay in posting this summary. There were a number 
of interesting questions and I did respond to most of the
individuals directly. Since many of the questions were
similar, the first part of this memo poses those common,
generic questions and answers them. This is followed by the
individual responses received. In the Appendix, a copy
of my original memo is included for reference.

I hope I have not missed any of the responses. If I did,
please let me know and I will post them.

There will be a panel discussion on the question "COULD
THERE BE REAL-TIME, INSTANTANEOUS LEARNING IN THE BRAIN?" 
at the World Congress on Computational Intelligence
(WCCI'98) in Anchorage, Alaska in May, 1998. I will post an
announcement on this soon.

The question of real-time, instantaneous learning is
tied to such other classical connectionist ideas as
local learning and memoryless learning. These three ideas
have led the development of various brain-like learning
algorithms for the last 40 to 50 years. So these open
discussions are about some of the most fundamental ideas of
this field. It is quite possible that we have been
developing the wrong kinds of algorithms all these years.

I hope more scholars from neuroscience, cognitive science
and artificial neural networks will participate in these
informal, open discussions and enrich this discussion. All
comments/questions are welcome.

Asim Roy
Arizona State University

------------------------------------------------
ANSWERS TO SOME TYPICAL, GENERIC QUESTIONS

(A) FIRST A CLARIFICATION OF THE NOTION OF "REAL-TIME,
INSTANTANEOUS" LEARNING

I think there was some confusion in my use of the term
"real-time." It should have been clear from my memo that I
was using the term "real-time" to refer to "Hebbian-style"
instantaneous and permanent learning. Hebbian-style
learning is used in such well-known algorithms as
back-propagation. In Hebbian-style learning, a training
example is used for some type of instantaneous adjustment
to the network and then the example is discarded
(so-called memoryless learning). So in Hebbian-style
learning, there is no recording or retention of any
particular information (a training example, that is) in
the system for subsequent use.

This is the style of learning generally used by all
learning algorithms in the field of artificial neural
networks. The key notion in Hebbian-style learning is that
of "instantaneous and permanent" learning from each and
every example provided. I used the term "real-time" to
refer to this mode of learning.

But "memory-based learning" can also be "real-time" in the
sense that learning can begin as soon as some information
about the problem is collected and stored. So, in that
sense, learning of motor skills in the Shadmehr and
Holcomb [1997] study could be considered real-time,
although it was not "real-time, instantaneous" in the
Hebbian-sense - the learning took about "5 to 6 hours" to
complete!! Learning in "5 to 6 hours" is real-time, but not
"instantaneous." If there is Hebbian-style "instantaneous"
learning in the brain, learning should have been "complete"
as soon as the practice session ended; it wouldn't have
taken any further time after practice and not the "5 to 6
hours" it took in this case.

I think most people understood my use of the term
"real-time" in the way I intended, but I realize that it
may have created confusion with some. "Hebbian-style
learning" instead of "real-time, instantaneous" would have
been a more accurate term.

--------------------
(B) A CLARIFICATION OF THE DIFFERENCE BETWEEN "MEMORY"
(SIMPLE RECORDING OF INFORMATION) AND "LEARNING"

Several comments (by Gary Cottrell, Gale Martin, Stefan
Schaal) result from confusion about the terms
"memory" and "learning". In some fields (and in everyday
life), the terms "memory" and "learning" are used
synonymously. So some of them claim that "learning" is
indeed instantaneous. However, they are actually
refering to simple recording of information that was
instantaneous, not to "learning" that was instantaneous.