Hebb (was PCT and Education)

CSGnet Archive CSG2005
1

[Martin Taylor 2005.01.11.10.13]

[From Bill Powers (2005.01.11.0632 MST)]

Increasing or decreasing a voltage in a neuron means nothing in itself with
respect to behavior. You have to know what function the neuron is
performing in the local circuit. Functions in general require hundreds of
connections, some excitatory and some inhibitory, adjusted to specific
synaptic "strengths", and not just the maximum possible strength.

Quite so.

The
Hebbian idea that more response from a neuron is better is just naive, as
is the vague concept of a "cell assembly."

Even in the 1950's, I never understood Hebb to think that "more is
better". And one does expect ideas to become a bit more refined after
half a century if they are basically good ideas. There are many
varieties of learning that we now call "Hebbian". For some, it is the
activity of the source of the signal that matters. For some it is the
activity of the recipient cell, and for others, it might be the joint
activity. I've forgotten which was Hebb's original, or which was his
final idea, but that isn't really relevant. The concept "Hebbian
learning" has come to mean "activity in a connection strengthens the
connection."

As to the vagueness of the concept of "cell assembly", remember that
underlying Hebb's "cell assemblies" would have been the "Pandemonium"
or the perceptron, each of which had very much the structure of the
presently presumed interconnections of the perceptual side of the
HPCT structure. If you call that naive, so be it. You are probably
correct, in the light of half-a-century of research.

At least nowadays we can say
"neural network," and we know that a lot more is required than just
assembling a bunch of cells and making them fire as fast as possible. Can
you imagine what would happen in the cell assembly we call a brain if all
the neurons exhibited the maximum possible degree of "Hebbian learning" and
began firing at the maximum possible rate? Instant death!

In the first quoted paragraph above, you mention inhibition. In this
paragraph, you forget it. I hope that was accidental, a function of
the unreasonably early hour of your writing!

It might be more instructive if you based your comments on the
results of modelling the effects of Hebbian-type learning in cellular
structures. I think it was Christian von (van?) der Malzberg that did
this in one or more articles in Kybernetika around 30 years ago, and
found that, for a multi-level visual system, what showed up were
various kinds of "cell assemblies" like those found by Hubel and
Wiesel in real brains, such as on-centre/off-surround receptive
fields, oriented line detectors, edge detectors, and so forth. I
don't remember the details or the reference, but I do remember being
impressed at the time that one could find such results by
computational modelling.

Von der Malzberg's (if I got the right person) work was done with
imposed stimuli (I think largely noise, but I can't be sure). But if
one were to redo the study in a control context, allowing the system
to influence its own input, the sensory data would become more
consistent, and one might expect that Hebbian learning in the
perceptual and the output sides would lead to refinement of the
control system -- better and more stable control with fewer
side-effects.

In my own work, I argued in a 1973 paper (The problem of stimulus
structure in the behavioural theory of perception. S. African. J.
Psychol., 1973, 3, 23-45) that Hebbian style learning (in the absence
of control) with lateral inhibition should cause the perceptual
apparatus to generate what amounts to a Principal Components analysis
of the sensory data (meaning that the sensory input would be
reconfigured into components that were statistically uncorrelated and
informationally maximally efficient).

The core of that argument was the effectiveness of inhibition. I
think the same must be true of von der Malzberg's computational
results, since he found that one of the low-level structure patterns
that developed was on-center/off-surround receptive fields and their
inverse. Such structures depend on inhibition.

Remember that in a Hebbian system, use would strenghten inhibitory
connections, as well as excitatory ones. It's quite true that if all
the connections were initially excitatory, the cell assembly would
self-destruct in a literal cloud of smoke. Organisms built that way
would not survive to propagate their species. But if inhibitory
connections were initially present in sufficient number (or with
sufficient breadth of effect), such an explosive catastrophe is not a
necessary result of Hebbian learning.

Maybe I could rediscover von der Malzberg's papers, as I think I got
reprints at the time. But maybe not. I'll send out a search party
(me:-).

Martin

2

[From Bill Powers (2005.01.11.1950 MST)]

Martin Taylor 2005.01.11.10.13 --

Even in the 1950's, I never understood Hebb to think that "more is
better".

That's what I got out of his proposals. From what I read, he thought -- he
said -- that learning was a matter of strengthening synaptic connections.
That's fine if you're thinking of qualitative input stimulus events leading
to qualitative response events as in classical conditioning, but it doesn't
work for quantitative circuits, particularly continuous control systems.

The concept "Hebbian learning" has come to mean "activity in a connection
strengthens the connection."

To me that says "more is better." I don't see how you can interpret it
otherwise. Why shouldn't activity in a connection weaken the connection? If
there's positive feedback, that is exactly what has to happen to eliminate
it. And for that matter, who says that learning is the same thing as making
synaptic connections "stronger," whatever that means? Is aiming an arrow at
a target just a matter of making all the synaptic connections, of either
sign, strong enough?

As to the vagueness of the concept of "cell assembly", remember that
underlying Hebb's "cell assemblies" would have been the "Pandemonium"
or the perceptron, each of which had very much the structure of the
presently presumed interconnections of the perceptual side of the
HPCT structure. If you call that naive, so be it. You are probably
correct, in the light of half-a-century of research.

I thought it was naive when I first read what he said, 50 years ago. Of
course there are cell assemblies in the brain. The whole brain is a cell
assembly, and it is made of smaller cell assemblies. That says nothing
about how they work, what they are for, or how they function together,
whether like perceptrons or anything else. It's arm-waving. It's like
explaining how a television set works by saying "It has circuits in it."

In the first quoted paragraph above, you mention inhibition. In this
paragraph, you forget it. I hope that was accidental, a function of
the unreasonably early hour of your writing!

It doesn't work with or without inhibition. Does the strength of an
inhibitory synapse increase with activity? If so, the brain will end up
with a collection of excitatory and inhibitory synapses, all at maximum
possible strength. This doesn't even come close to explaining what really
needs explaining. We have to explain how weights in active synapses are
adjusted both upward and downward, for both excitatory and inhibitory
synapses, so as to produce just the circuitry that will extract perceptions
from lower-order signals, create comparators and connect them to output
functions, and so on. I say that the idea that this can happen simply as a
result of active synapses gaining strength is totally inadequate. Making
synapses stronger isn't what has to happen. I've never liked Hebbian
learning since I heard of it, which was a long time ago. It didn't meet any
of the requirements of a model.

It might be more instructive if you based your comments on the
results of modelling the effects of Hebbian-type learning in cellular
structures. I think it was Christian von (van?) der Malzberg that did
this in one or more articles in Kybernetika around 30 years ago, and
found that, for a multi-level visual system, what showed up were
various kinds of "cell assemblies" like those found by Hubel and
Wiesel in real brains, such as on-centre/off-surround receptive
fields, oriented line detectors, edge detectors, and so forth. I
don't remember the details or the reference, but I do remember being
impressed at the time that one could find such results by
computational modelling.

I'll bet his model contained a lot more rules than "activity strengthens
synapses."

The core of that argument was the effectiveness of inhibition. I
think the same must be true of von der Malzberg's computational
results, since he found that one of the low-level structure patterns
that developed was on-center/off-surround receptive fields and their
inverse. Such structures depend on inhibition.

Did the amount of inhibition in either paper increase with activity in the
inhibitory synapses? If so, you'd need to add some ad-hoc external system
to prevent runaway, for in mutually inhibiting systems there is a very
precise limit on the strength of the inhibitory effect. At that limit,
positive feedback just becomes unity, and for any larger amount runaway occurs.

Remember that in a Hebbian system, use would strenghten inhibitory
connections, as well as excitatory ones.

That wouldn't help, unless you imposed some very special design on the
system to rule out having an even number of inhibitory synapses in any
given closed loop. Doing that, of course, would show that so-called Hebbian
learning does not work in general, but requires very special conditions
even to avoid disaster, much less accomplish anything positive.

Maybe I could rediscover von der Malzberg's papers, as I think I got
reprints at the time. But maybe not. I'll send out a search party
(me:-).

That's a good idea.

Best,

Bill P.

3

[Martin Taylor 2005.01.11.23.04]

[From Bill Powers (2005.01.11.1950 MST)]

Martin Taylor 2005.01.11.10.13 --

Even in the 1950's, I never understood Hebb to think that "more is
better".

That's what I got out of his proposals. From what I read, he thought -- he
said -- that learning was a matter of strengthening synaptic connections.

I can't remember so far back as when I read Hebb in the original, but
I do know that he was a rather subtle thinker, and most highly
thought of by my first boss, who knew him well and was himself one of
the most intelligent people I ever met.

I'm quite ready to be corrected on this, but I cannot imagine Hebb
having failed to include a concept of the decay of synaptic
connections in the absence of activity. The effects of positive
feedback were not unknown, even then, and I seem to remember that
Hebb, like so many of the well-known psychologists of that era,
started as an engineer. I also know that in the old days when I was
concerned with this kind of thing, the decay of little-used synapses
was one of the aspects to be taken into account.

I think it was Christian von (van?) der Malzberg that did
this in one or more articles in Kybernetika around 30 years ago, and
found that, for a multi-level visual system, what showed up were
various kinds of "cell assemblies" like those found by Hubel and
Wiesel in real brains, such as on-centre/off-surround receptive
fields, oriented line detectors, edge detectors, and so forth. I
don't remember the details or the reference, but I do remember being
impressed at the time that one could find such results by
computational modelling.

I'll bet his model contained a lot more rules than "activity strengthens
synapses."

Of course. But when his work is referenced, it is called "Hebbian".

Maybe I could rediscover von der Malzberg's papers, as I think I got
reprints at the time. But maybe not. I'll send out a search party
(me:-).

That's a good idea.

I tried, today. My paper reprints are not organized as they should be
(look at my picture on my web site, and understand that I have had
two office moves since then, and you'll understand why). Our library
has most of its journals packed away in boxes in the attic because of
building renovations, so I will have to find time to go downtown and
look for it in the University of Toronto library. If it wasn't von
der Malsburg I'm thinking of, the relevant papers won't be far away
from his.

Through a Web search, I managed to find that my memory for his name
was as accurate as my memory for names usually is. His real name is
Christof von der Malsburg, and he is now at Universitat Bochum and at
U Southern California. I'm not at all sure, though, that he is the
author of the papers I am thinking of. In the 70's he was working in
the area of self-organized visual systems and he did have a relevant
paper in Kybernetik, which I could not get other than as a citation.
It's about self-organized oriented line detectors (Kybernetik, 1973,
14, 85-100).

···

-----------------

I think that what I find disturbing in this discussion is that you
give the impression of wanting to produce a straw man that you can
simply blow down. You lift a single salient element out of a theory
that was well enough thought out to impress quite a few people who
were willing to subject it to different tests. You then show that it
takes almost no thought to see that this element by itself cannot do
the job that those people thought the theory could do. You imply that
all these people were so incapable of even a little thought, that
they spent years working on something you could demolish in only a
matter of minutes. Isn't that just a touch arrogant?

For myself, I would prefer to think that if a theory were so easily
demolished, and yet so many clever and qualified people took it
seriously, I might have missed something. Perhaps I could have seen
something they all overlooked, but I would think it more likely I had
overlooked something they had all seen.

Martin

4

[From Bill Powers (2005.01.12.0622 MST)]

Martin Taylor 2005.01.11.23.04 –

I can’t remember so far back as
when I read Hebb in the original, but

I do know that he was a rather subtle thinker, and most highly

thought of by my first boss, who knew him well and was himself one
of

the most intelligent people I ever met.

I wasn’t impressed the same way – but testimonials like these (or
opinions like mine) are not, thank goodness, the criterion by which we
judge the worth of ideas in science. I’m just saying how his work struck
me. Maybe a deeper acquaintance would have changed my mind, though as I
now look through some books of his that I own, I doubt it.

I’m quite ready to be corrected
on this, but I cannot imagine Hebb

having failed to include a concept of the decay of synaptic

connections in the absence of activity.

Fine, that would make them into leaky integrators of a sort (gain, not
signal, integrators, though). I just can’t believe that
“activity” alone is enough of a basis for reorganization. Have
you actually modeled a neuron that works this way?

Bjorn Simonsen’s quote from Hebb (via Wikipedia) is worth
requoting:

Let us assume that the
persistence or repetition of a reverberatory activity

(or “trace”) tends to induce lasting cellular changes that add
to its

stability… When an axon of cell A is near enough to excite a cell B
and

repeatedly or persistently takes part in firing it, some growth process
or

metabolic change takes place in one or both cells such that A’s
efficiency,

as one of the cells firing B, is increased. "

That sounds interesting on first reading, but what it says, when you
think about it, doesn’t make any sense. It really says that connections
are made, and strengthened, at random, because WHATEVER cell the axon of
cell A is near gets a connection made to it, regardless of function or
consequences. This is nowhere near enough to assure that a control system
will appear, or any particular form of organized system. What is missing
is a system that monitors the consequences and on that basis causes
connections to be made, modified, or broken independently of happenstance
proximities. In other words, the formation of a connection can’t be
determined totally by local conditions if the resulting organization is
to have global usefulness.
You can patch Hebb’s idea up by adding spontaneous decay of connections
and selected effects of inhibition to prevent runaway, but the result
still does not explain how local conditions can be sufficient to
determine useful global functions. I think the basic approach is doomed.
It can’t explain how a pigeon can learn to walk in alternating left and
right circles to get fed. The walking neurons have nothing to do with the
eating neurons, yet the state of the latter has to affect the connections
to the former.
Hebb’s adopted concept of drive reduction was a good start toward control
theory – if followed through it would lead to a control model. The
problem is that he failed to differentiate between drives that lead to
behavior and drives that lead to learning, the first being (in PCT)
hierarchical error signals, and the second being intrinsic error signals.
This led him to great confusion about the role of drives:
“These are conditions in which we can understand how drive reduction
strengthens one form of behavior and abolishes others. But the
explanation applies only when drive level is high enough to produce some
degree of cortical disorganization.” (A Textbook of Psychology
(1958); p. 141)
You see the problem. We have here a mixture of an ordinary error signal
producing the behavior that reduces it, and the same error signal,
at a higher intensity, producing reorganization. Of course we have our
own version of this same observation, but we distinguish between ordinary
and reorganization-causing error signals – in general they are signals
in *functionally different systems (*if not physically different).
So no explanation based on purely local effects can work if our theory is
correct. I probably had exactly this problem in mind when I wrote that
reorganization is something done to system, not by a system
(to itself). In other words, I was rejecting the concept of a
self-organizing system produced by completely localized
effects.
If Hebb had not tried to explain reorganization as a passive consequence
of proximities, he could easily have arrived at control theory complete
with a reorganizing system. But once you have a basic concept that seems
to work, there is an almost irresistible temptation to keep patching up
problems as they appear instead of starting over by looking at the root
idea each time a problem shows up. When you go only a little distance in
this direction, you start making your idea work instead of just
letting it develop.
On the cover of Hebb’s Organization of Behavior, the subtitle is
very clear about the initial idea that Hebb adopted, and then kept trying
to patch up:

“Stimulus and response – and what occurs in the brain in the
interval between them.” My copy is a reprint of the first edition
from 1961 – one year after

Clark, MacFarland, and I had published our first two-part paper on
control theory, and after it had been reprinted in several places (such
as the Journal of General Systems Theory). So Hebb was proclaiming
himself to be a proponent of a basic concept in which I had ceased to
believe long before 1961.

I think that what I find
disturbing in this discussion is that you

give the impression of wanting to produce a straw man that you can

simply blow down. You lift a single salient element out of a theory

that was well enough thought out to impress quite a few people who

were willing to subject it to different tests. You then show that
it

takes almost no thought to see that this element by itself cannot
do

the job that those people thought the theory could do. You imply
that

all these people were so incapable of even a little thought, that

they spent years working on something you could demolish in only a

matter of minutes. Isn’t that just a touch arrogant?

Maybe, maybe not. It’s what I think, and as I said I’m not impressed by
testimonials. It is perfectly possible for highly intelligent people to
have a wrong idea and support it, even though later it seems perfectly
obvious that the idea was wrong or even silly – in the light of other,
newer, ideas. Most scientists believed in phlogiston, and could prove it
existed, over a period of about 150 years. There must have been a time
when most of the great thinkers who lived believed the world was flat,
and could prove it. These were not stupid people, but not being stupid
didn’t keep them from being wrong.

Anyway, I didn’t reject Hebb’s ideas over a period of minutes, but over
several years. I think he was wrong to believe in stimulus-response
theory (he should have known about cybernetic concepts, which do not
appear in the indices of either of the two old books of his that I own).
I think that his concept of cell assemblies was vague, and his use of
“reverberating circuits” or “closed pathways” as the
mechanism of memory was flat wrong (regardless of how venerable that idea
was even then). He recognized “the presence of closed pathways, or
loop circuits, which are found throughout the nervous system.”
(Textbook, p. 103), but because he picked the wrong role for them he
missed his chance to rediscover control theory.

For myself, I would prefer to
think that if a theory were so easily

demolished, and yet so many clever and qualified people took it

seriously, I might have missed something.

Come on, Martin, you know that’s no argument all. Ideas are not true or
plausible just because clever and qualified people believe in them. Who
says they’re clever and qualified anyhow? Only other clever and qualified
people who think the others and they themselves are right. Just
think of all the clever and qualified people who have believed in S-R
theory, and all the other clever and qualified people who have
disbelieved it. Somebody gave them both PhDs.

Perhaps I could have seen
something they all overlooked, but I would think it more likely I had
overlooked something they had all seen.

That is a properly humble and non-arrogant attitude, unless of course you
are defending a prior committment to some of these ideas which now are
being called into question. You do, after all, cite a published paper of
yours in which you took Hebb’s ideas seriously. To defend Hebb now is to
defend your acceptance of his ideas, as well as whatever you built on
them, which is something of a conflict of interest.
Anyhow, I’d rather see a defense of Hebb based on his ideas, not on the
reputations of those who have supported them.
Best,
Bill P.
P.S. By the way, I also have the fourth edition of the Textbook of
Psychology (1987), this one by Hebb and Donderi. On the flyleaf is
hand-written

"To Bill Powers -

with appreciation for the force and clarity of your feedback
idea.

Don Donderi

88/1"

I’m sure I had invited him to join the CSG, which was then three years
old, but nothing came of it. Wonder what he would say now. The book is
still pretty pure S-R theory, even in 1987.

5

[From Bruce Gregory (2005.0112.1027)]

Bill Powers (2005.01.12.0622 MST)

But once you have a basic concept that seems to work, there is an almost irresistible temptation to keep patching up problems as they appear instead of starting over by looking at the root idea each time a problem shows up. When you go only a little distance in this direction, you start making your idea work instead of just letting it develop.

Amen.

The enemy of truth is not error. The enemy of truth is certainty.