🚀 go-pugleaf

In article <343AAFF1.6D35@dial.pipex.com>, Howard Snart
<aac88@dial.pipex.com> wrote:

>[...]
>This approach stems from several misgivings I have over the existing
>methods of teaching neural nets... most significantly
>backtracking.(which I see as un-scalable, and fundementally wrong... you
>force the outputs to conform, whereas the problem may lie far from the
>outputs)

  I agree.  Traditional neural nets are not biologically plausible,
nor do they do the kind of learning that is needed to gain common
sense.

>I also think that we know all we need to know to produce a good
>description of how the brain functions.

  I think we're getting close to something big.

>I base this scheme on the following (established) points:

  Established by whom?

>1-firing a neuron strengthens its links

  I have to disagree with this.  Empirical evidence from
neurophysiology (I'll see if I can dig up the references for this
later) and our own model indicate that unless a synaptic potential is
strongly associated with a certain temporal order, the firing of a
neuron will always induce a weakening of its downstream links, not a
strengthening.  I know this is counter-intuitive but our results
leaves no room for compromise here.  There is only one input condition
that causes a synapse to get stronger.  This is related to temporal
order.  We'll have more to say about this later.

>2-a neuron firing causing another one to fire strengthens the link even
>more.

  I agree and I would add that this (causal order) is the only thing
that strengthens a link or synapse.

>2a-a neuron firing, without an input firing, weakens that input link.

  Yes.  Note that the link is strengthened only if it was already
firing when the neuron fires.  It is weakened otherwise, even if its
starts firing after the neuron begins firing.  There is strong
evidence for this.

>3-the links strength slowly degrades through lack of use, and can
>eventually be destroyed.

  This seems to happen in natural neurons as part of the observed
random firing of neurons.  We believe it's the mechanism for
forgetting.

>4-once fired, the link is temperorerily boosted(its weighting is
>increased), this decays quite fast. I term this increase 'boost', for
>want of a better term.

  I'm not sure what you mean by the fast decay part.  Again here, I
would disagree that the simple firing of a neuron is enough to
strengthen its synapses.  The neurophysiological evidence says
otherwise.  We interpret this as part of the natural method used by
the nervous system to weed out non-associated synapses.

>5-a neuron has a 'recharge' time, thus they cannot fire continusouly.
>They can only fire, typically, once every millisecond

  We believe that spiking frequency is has to do with two things that
have only an indirect relation to neural information processing.  a)
It is an indication of temporal resolution of a neuron.  b) The higher
the frequency, the stronger the signal (the amplitude of the spikes
remains constant).  A strong signal is important for specific
downstream synaptic strengthening purposes.  What is more important to
us is spike train initiation (ON) and cessation (OFF).  We get
excellent results by dealing only with these two events.

>ok,
>now, given a weakly connected, deep, network... my equivalent to a new
>brain..

  In our models we start with a tabula rasa but it's almost the same
thing as a pre-connected network in which the synapses have zero or
little strength.  The depth (number of layers) depends on 3 factors,
a) the machine's performance characteristics, b) what is learned, and
c) the designer's goals.

>These neurones are connected like those in a biological brain, and
>signal propogation along the synapses is not instantaneous. i.e. it
>takes time for a signal to pass down the synapse.
>...and a source of input... say an eye.. i.e.

  Signal propagation speed is important but not crucial.  What is
crucial is the temporal relationship between signals.

>        |---------------|
>eye --> |    brain      | --> outputs
>        |---------------|
>
>Now, when I stimulate my 'eye':
> 6-from 1: The input neural firing will not initially propogate, because
>of the weak synapses. But, this firing strenghtens the input synapses,
>and, eventually, the next layer will fire. Thus it progresses, so
>eventually, given enough stimulous, we will end up with outputs as a
>result of inputs... or be they random.

  OK.

> 7-from 3: If it does not get input, then a neurons strength will
>degrade, and could eventually be destroyed. 

  OK.

>This next point could be demonstrated with just my 'eye', but, for
>clarification, i'll include another input source... an 'ear'
>
>i.e.
>ear --> |---------------|
>eye --> |    brain      | --> outputs
>        |---------------|
>
> 8- from 2,6: Now, because I am being stimulated by the 'real world', my
>inputs will be corralated. e.g. if I sit watching someone clap their
>hands. Thus i will have waves of neural firings propogate from my input
>neurones. Thus I will have constructive interference where those waves
>intersect. i.e. The synchronous inputs will cause those neurons
>stimulated as a result of both inputs to be reinforced.

  This is the straight Hebbian adaptive mechanism.  Unless a temporal
factor is added, not much will be learned by this.

> 9- from 8: Thus, I should see structures emerge that correspond to the
>corralation of the inputs. 

  Sure.  But you need a strong causal ordering mechanism for
correlations to be useful or even meaningful.  Actions (output effects
or behavior) will depend on conditions.  Conditions are the sufficient
causes of effects.  Unless the network is designed to deal with causes
and effects from the start, not much will be learned.

> 10- from 9: Now, given corralated input, these corralation points
>should be adequately stimulated, and be strong centres of neural firing.

  OK.

> 11- from 10: Thus, I should end up with outputs that correspond to
>corralated inputs. e.g. by looking at the outputs I might be able to
>find a neurone that fires when I am looking at a red bell. 

  Maybe.  But I doubt it.

>Note, now we can corralate inputs and produce some outputs based on that
>corralation. This is still not very usefull as the outputs are random.
>If I add a further 'sense'. This equates to 'right' and 'wrong'. It
>fires more when the current situation is 'good'. 

  Hmm.  'Sense' is vague.  This is getting into unchartered territory.
IMO, "good or bad" is decided at the single neuron level.  In a causal
system, there will be conflicts that will cause connections at
multiple competing neurons to simultaneously weaken.  The severing of
one or more connections usually ends the conflicts.  The winners are
the strongest connections.

>i.e.
>ear --> |---------------|
>eye --> |    brain      | --> outputs
>good--> |---------------|
>
> 12- from 11: I should now produce outputs that tend to corralate, and
>maximise my inputs. Outputs that cause my good sense to fire more will
>increase the good sense, and therefore become prevalent. 'bad' behaviour
>is not removed, it is superceeded by 'good' behaviour.

  This is getting into even more uncharted territory.

>But, what is 'good'?, and how does it apply?
>Now, if I remove the good sense, and replace it with a basic set of
>drives. This equates to a very basic, low level, behaviour. This equates
>to our basic nervious system. 
>e.g.
>
>ear --> |---|---------------|---|
>eye --> |-A-|    brain      |-B-| --> outputs
>        |---|---------------|---|
>
>A = input basic nervous system. 
>B = output basic nervous system.
>
>The job of A is to simplify the environement, to amplify 'good' input(
>or reduce 'bad' input). 
>The job of B is to simplify the brain outputs, and to provide multiple
>paths to the actual outputs.
>
>The initial basic drive could be 'closeness to parent = good'. And, in
>my simple creature, this could be represented by my 2 sense organs as
>'red' and 'noisy'.
>The basic nervous system, upon encountering 'red' or 'noisy' will
>amplify that data, or reduce input that is not 'red' or 'noisy'.
>
> 13- from 12: 'good' imputs, i.e. the strong inputs, are likely to
>produce the initial structures, and will stimulate the brain more. 
> 14- from 13,6,8: Thus 'good', corralated behaviour is likely to result.
>Outputs that enhance 'good' behavoiur will be prevalent, the net will
>have learned to solve the situation encoded in A and B.

  This is what David Longley would call "intensional" stuff and I
would agree.  The best thing to do in situations where speculation is
beginning to get out of control is to run a set of experiments.

>Further:
>15- from 3: Old memories will gradually degrade, if not used. The
>destructuion of unused, or redundant, neurones could be seen as an
>optimisation process.

  OK.

>16- from 4: Using a thought, gears up the mind for thinking that way.
>So, for example, the brain can gear itself up to spotting patterns, like
>seeing a computer game when your eyes are closed. 

  I have no idea what this means.

>17- from 4: This is equavalent to the current 'thought' that the brain
>is having. And is equivalant to an idea 'mesh'.

  The only way you are going to have anything related to something as
vague as "current thoughts" is to have an attention focusing
mechanism.  We have a subsystem that we call the "searcher" that lives
in the network.  It can perform this function.

>19- from 17,4: If the neurone becomes 'stressed' it could be that its
>'recharge' time, and boost duration become large. This would make the
>brain stagnate, and reduce its capacity for new thought ( the current
>thought links would be too established). Thus sleep is needed, to
>replenish the neurones and reverse this process. 

 OK.  Biological neurons may need this but not simulated neurons.

>20- from 19: It is possible that this process is important, so that an
>important idea can dominate, so a solution can be found.

  OK.  We believe that attention focusing is indeed important but only
in large systems and focusing on the right stuff can be learned
through conditioning.

>21- from 19,20: 'stressed' neurones = overloaded boosted links =
>physical(chemical) exhaustion. The increased long term link between
>neurones means gets stressed less, under the same load.

  I don't know what this means.

>22- from 8,9,10,11: If there is a structure that is regularly
>stimulated, like 'red'. Such structures could be regarded as inputs to
>the rest of the brain. Thus corralation between these can prevail. Thus
>complex, derrived, behavoiur can be produced.

  I don't know what this means.  The same association mechanism should
take care of all correlations.  Our model does not actively look for
associations.  It uses a processes of elimination to weed out
non-associated synapses.

>23- from 5: This prevents cyclic loops on the small scale, and so helps
>prevent paralysis.

  I don't know what this means.  Have you conducted any experiment and
observed these "cyclic loops?"

>NB:
>-this scheme will not work well with small numbers of neurones.
>-there are a lot of variables to this scheme, and finding values that
>promote convergence might be hard.
>
>Sorry to go on, but i'd be very inturested to know what you think.
>This represents a condensed, and simplified version of the scheme. There
>are further concequences which I havnt mentioned here.
>
>Howard Snart

  OK.  What you have described is mostly the traditional Hebbian form
of learning.  You are onto something IMO, but you need to think *cause
and effect*.  This is the main key to the puzzle.  We'll have more to
say about the role that cause and effect play in the nervous system
and about how this role fits into the behavior and meaning of synapses
in the near future.  Synapses do not all behave the same way.  For
now, I'll just say that in addition to the complementary YES
(excitatory) and NO (inhibitory) forms, there are two other
complementary types of synapses.  More on this later.

Regards,

Louis Savain

President, Marengo Media, Inc.