On Fri, 22 Sep 95 09:02:54 EDT,
CL1779A at american.edu (Christopher T. Lovelace) wrote:
>As to what neural mechanisms underly these properties, I'm not sure. Perhaps
>a type of Hebbian learning? When I learn a song, I often listen to a
>recording over and over. This repetition leads to consolidation into long-
>term memory (hippocampus). The memory itself, I would imagine, is stored in
>cortex, in a fairly distributed fashion.
Sorry but I must disagree with you because I have done a lot of research
on this and came up with different results. The current view (proposed
by several books on the subject) that long term memory is stored in the
hypocampus is completely wrong.
Our brains is working on a pulse coded system. These pulses flow in
"strings". The whole mechanism is based on the concept of free
association. That is the knowledge that two things are related without
having to know HOW they are related. Our brain is wiring together events
and things in the order in which we perceive them either by direct
observation or by thinking about them.
Like this the brain is building strings of events and things that can be
regarded as some kind of story (but it can just the same be a sequence of
muscle actions to fire).
If we see, hear, feel or smell something the first node of a string is
activated. Because humans have developed in to a "smart" species that
can think about events when they are long gone we need a "device" to fire
pulses in those strings. A string can be activated by an external event
(e.g.smell) but to activate the complete string and all related strings
connected behind it that draw pulses from the first one a lot more pulses
are needed that the external event will provide. The hypocalamus is
doing this. The amount of pulses put out by the hypocalamus is more or
less constant over short periods so that the amount of strings that can
be "kept alive" is also more or less constant. If a string (thought) is
finished and none other are active then the hypocalamus (we call it pulse
source in AIR1NN (which is a revised kind of NN that mimics the behavior
of biological neurons)) will activate new strings almost at random
(mostly already partially activated by being related to the previous
thoughts). This is why your thoughts never stop. If an external event
is activating new strings then this new string will "suck" on the
connections to the pulse source so that the pulse flow is redirected to
the newly activated strings. If there are to many new strings activated
or they draw to much pulses (several strings starting from the newly
activated one) then the stream of pulses to the previously active strings
will stop completely and the person will "forget" what he was doing
before and continue with the new stream of thoughts.
This is an oversimplification but our memory works by association. In
other words it will put out anything that is related to what is put in
into the network of strings. In the example of driving a car each next
action to take is related to the place where that action took place
earlier. This is a very very powerful retrieval system but it also has
one great disadvantage. You already need to know at least a part of what
you want to remember. There are also other problems that could be more
easily solved in software then in a biological unit (human/animal). So
the future is to the computers (Don't shoot, I don't like it either).
If the hypocampus is damaged the amount of pulses that is needed to train
our biological neural net (each pulse is one training) can not be
achieved for "internal" strings. Internal are all strings that are not
almost directly connected to external events (senses). So without
hypocalamus you can still learn new motoric skills. Forming abstract
strings require the pulse source to provide the amount of training needed
to form new memories (alter synaptic connection strengths).
The hypocampus plays also a very important role in sleep and our response
to chemical stimuli. Pain can (even when more local and imprecise)
replace the hypocampus for learning because pain is nothing else then an
enormous external pulse source. So pain will increase the short term
learning skills dramatically. In the long run it is a bad alternative
since this amount of pulses is to powerful to get the strengths in the
neural network right. The network will "over-learn" the activated
strings. The same is happening during "normal" learning (when awake) so
that we require regular sleeping periods during which the pulse source is
reactivating the strongest connections (over-learned strings) because
these are at that moment the strings with the smallest "resistance"
(strongest neural connections). These strings are then supplied with a
small amount (30% of awake) of pulses. This amount is not enough to get
the entire string activated (charges decay after a few seconds) so that
the charges (pulses) are spreading more evenly to all connected other
strings. Like this connections (other then the over-learned ones) will
also fire so that those will become better and the over-learned a bit
weaker.
To understand how this is working I must first explain that the output of
a neuron is also an input. A neuron will fire if the charge in the
front-chamber (dendrites and the "body" of the cell) exceeds a threshold
compared to the back-chamber (axon). The potential of the axon is the
potential of its environment. If a second neuron (which dendrite is
connected to the axon of the first one) is firing then that second neuron
will have a negative charge in its front-chamber after firing. This in
turn will cause the charge of the axon of the first neuron to become less
so that the first neuron will fire easier as long as that condition
exists. This effect is causing some kind of "vacuum" that is sucking
charges from previous neurons. The longer the string of firing neurons
formed this way the stronger this effect becomes (that is also how the
pulses from the pulse source end up connected to those strings). However
this only takes place if the next neuron fires. If you keep the amount
of pulses over time low then you will not activate many of the next
neurons in the string(s) because the charge decays before it can build
up.
So in practice the distribution of pulses from a neuron to the next ones
(all other neurons who's inputs connect to the output of this one) is
determined by:
1) The neural weights of the connections. The better (stronger) the
connection the easier that next neuron will fire.
2) If one of the next neurons has fired shortly ago then the connection
to that neuron will seem to be better for a short time because of the
negative charge in the front chamber of that particular next neuron.
Because of (2) a string will form. In other words: The pulses will chose
a path and stick to it until forced otherwise. In a "normal" neural
network (AI software) all of the next neurons get some of the value of
the previous. In our brain this is not happening. A few (mostly only
one) of the next neurons get all of the pulses and the remaining non at
all. Current NN software only mimics behavior (1) and not behavior (2).
It only works by calculating distribution of charges (values) but that is
not what matters. What matters is the path that the pulses are choosing
(what I call a string). However if (2) happens then (1) will change to
adapt to the behavior of (2). So if a path is formed then the strength
of the neural connections will change in favor of those that form the
path of the pulses. This effect can not be turned of in our biological
neural nets so if we repeat a movement or thought very often then it will
become over-learned. So the next night we will dream of those
over-learned movements or thoughts (yes AIR1NN needs dreaming). Before
this "leveling" of neural connections (during sleep) can start all
activity has to stop first because if there are still strings active then
the pulse source will feed those (many pulses to a few strings) instead
of providing wide-spread low-level activation (few pulses to many
strings). After a while (can not happen fast because charge decay times
are same as awake) those over-learned connections will become less
good/strong (strong connection=small resistance). Once this has happened
there will be no favor for particular strings to receive pulses from the
pulse source (hypo calamus) so that normal operation of the network
resumes. That means that strings start to activate again in sequence in
the order in which they are connected together forming thoughts similar
to and in the order of those real experiences. While that is taking
place the pulse source if activating less strings with more pulses until
the level of activation is reached and effect (2) star ts to work again.
And then we wake up.
Of course there is a lot of chemical influence on the hypocalamus to
force start and end of sleep but it all fits nicely together.
- awake(over-learning)
(pulse source at 100% with variations from
50 to 200% (rest to nervous/aggressive))
- first-sleep(period of non-activity)
(pulse source at 5%)
- first-dreaming(dreaming about things done that day)
(pulse source slowly 5->30%)
- REM-dreaming(thoughts become more story-like)
(pulse source at 30% increasing -> 50% to wake up)
So as you can see theory and observations match nicely.
This is becoming a long story and there is a lot more to tell that is not
related to the question of the original poster so if there are any more
questions, things that you think are wrong or comments... feel free to
send me a mail. If you wish I can also mail you a copy of earlier
articles and postings on this subject.
With kind regards,
-----------AIR1-research-----------
EMAIL TO: Stephan.Verbeeck at ping.be ---------------------
ADDRESS : Helderbeekstraat 41 Life is but a dream
3582 Koersel (Belgium) ---------------------
TEL : 011/42.16.46
------------------------------------
