🚀 go-pugleaf

In article <34300f6c.0@news.ttlc.net>, <modlin@ttlc.net> wrote:

>
>Louis,
>
>Perhaps you can help me to understand something?
>
>I looked up the references you gave, and it does appear that there
>is enough substance to the evidence for "differentiial" Hebbian
>mechanisms (reinforcing only inputs which precede the response
>in time) that I can't dismiss the notion out of hand.
>
>However, that leaves me confused as to how to address a problem
>that I came across several years ago, and thought I had resolved by
>accepting that the basic Hebbian mechanisms were not sensitive to
>the ordering of stimulus and response, that correlation rather than
>causation was indeed reinforced.

  This is what I call a straight Hebbian coincidence learning
mechanism.  Not much can be learned by a network using this approach.
Our model uses a process of elimination to weed out uncorrelated
(unassociated) inputs.  Only temporally associated inputs survive.
The important question is, associated to what?  We'll publish the
answer in our coming announcement.  Please be patient, Murphy and his
pesky laws are conspiring against us one more time.

>If a hebbian-adaptive cell learns to recognize some cluster of events
>in a continuing ensemble of inputs, it will necessarily reinforce
>any correllated predecessors to the events which it is currently
>recognizing.

  True.  A straight Hebbian cell will also reinforce correlated
successors who happen to arrive while the cell is firing.  This
conventional Hebbian approach does not differentiate between
successors and predecessors.  But, given your example in which
predecessors are favored over successors and coincident signals, a
different behavior is bound to emerge.

>As the earlier-occurring signals acquire weight, the cell temds to fire
>earlier, shifting the center of its response toward the leading edge
>of a sequence of correllated events.
>
>If it does not also reinforce inputs arriving after its current firing
>time, the originally-learned inputs lose weight as it moves forward.

  This is true if you are using the conventional method of input
summation or integration to determine when the cell should fire.  As
we will show in our coming presentation, this integration method,
known as the threshold method, is fallacious.  There is a precise
input firing pattern that determines when a cell fires and how long it
continues to fire.  This is related to causation.  I regret that I
must leave the rope dangling at this juncture.

>Unfortunately, this leads to an intractible instablility, in that
>the response will drift to the leading edge of a correllated sequence,
>where it will hover until transient correlations entice the cell into
>firing a bit earlier, reacting to events not reliably correllated with
>the original series over a longer term.

  I agree that this would happen given the conditions you imposed.

>Since the original adaptation has been "forgotten" (due to asymetric
>reinforcement), the cell does not acquire a stable response... it keeps
>moving on to different things.

  Correct.

>This turned out to be (at least so far as I could determine) an inherent
>weakness in the assumption of precursor-only reinforcement, which could
>not be overcome by "tweaking" any other aspects of the model.

  I think you are overlooking something important and making one or
more wrong assumptions.  Sorry.

>I could fix it only by giving up on asymetric reinforcement, allowing
>reinforcement to extend to inputs occurring both before and after the
>post-synaptic response.

  This is the straight Hebbian approach of input strengthening within
a given integration window.  We are 100% convinced that it is
incorrect.

>As it turned out, I found this to be a useful step in a number of ways,
>and discovered that I could re-introduce the learning of causality at
>higher levels in a much more robust fashion.  So I was, until now,
>quite happy with the assumption of symetric reinforcement.

  I can see why.  Given the assumptions you made, I would have come to
the same conclusions.

>But if that is truly not the way real cells work, I ought to
>understand how that can be.

  You just put your finger on it.  Cells don't work that way.  The
differential Hebbian mechanism of synaptic modification is correct but
it is not the whole story.  If it were, the instability problem you
identified would be insurmountable.  We think we know what the whole
story is and you can rest assured that its simplicity will surprise
most of you who are currently doing research work in AI and/or
neurobiology.

>So.  Can you tell me how to avoid the instability associated with
>anticipatory drift, which appears to me to be an unavoidable result
>of the model you are espousing?

  It has to do with your method of integration.  If a biological
neuron is this meticulous as to its method (temporal order) of
synaptic reinforcement, don't you think that its method of input
integration would be just as meticulous?  The solution to the problem
raised in your pertinent post will be revealed as soon as our ANML
package is ready.  ANML stands for Associative Network for Machine
Learning.  Thanks for your patience and interest in this.  We think
this is the most exciting time in intelligence research in decades.

Regards,

Louis Savain

President, Marengo Media, Inc.