Re: Symbolic AI: Why Marvin Minsky and Curt Welch Are Out to Lunch

Traveler <traveler@xxxxxxxxxx> wrote:
On 22 Jul 2006 21:14:23 GMT, curt@xxxxxxxx (Curt Welch) wrote:

Traveler <traveler@xxxxxxxxxx> wrote:
On 22 Jul 2006 12:32:45 -0700, "JGCASEY" <jgkjcasey@xxxxxxxxxxxx>

How can we use those broad statements to implement some kind
of intelligent machine? That is, how can we convert those views into
software or hardware that demonstrates what you mean exactly?

1. One can start with sensory processing. The realization that the
meaning or identity of a signal in a stream depends on its temporal
positions with respect to signals in other streams is essential. It
follows that sensory cortex neurons are signal separation processors.
They separate specific signals from a given stream and sends them down
a specific path.

What do you see being the correct signal separation algorithm? And
what's the basis of your belief that the algorithm you use is the
correct one?

The basis of my hypothesis is at once religious, philosophical,
scientific and experimental. I will not go into my religious views
here since this is not the proper forum for it. Phlosophically, I
approach all problems (scientific or otherwise) from a yin-yang
complementarity POV. Examples of complements are sensors/effectors,
event onset/event offset, etc... The scientific basis of my approach
to signal separation comes from a paper written by Henry Markram et al
in 1997. It is titled "Regulation of Synaptic Efficacy by Coincidence
of Postsynaptic APs and EPSPs".

http://www.nbb.cornell.edu/neurobio/linster/BioNB420/pdfs/markram_etal_19
97.pdf

Essentially, Markram found that the efficacy of synapses impinging on
a pyramidal neuron in the sensory cortex is strongly increased if they
fire about 10 ms before the target neuron fires. If they fire at any
other time, they are slightly weakened.

So you are saying if the input fires 10 ms before the output fires right?

Discrete signal separation is not rocket science. It is a simple
probabilistic process. Although Markram does not explain in his paper
how the neurons are triggered to fire in vivo (he conducted in vitro
experiments), I hypothesize that a sensory cortex neuron fires only if
one of its input synapses (the successor) fires immediately after
(within 10 ms in the human brain) a predecessor synapse.

You have lost me. Are you talking about two different synapse (inputs) to
a single neuron? And the relation they fire relative to each other? So
you aer saying that the neurons fire if two inputs fire within 10 ms of
each other?

The neuron
works as a gate or filter that opens to let some signals through while
blocking the others. Its function is simple: a specific signal within
a sensory stream is channelled down a specific path.

In my own experiments, I have found that a 10 to 1 correlation ratio
between predecessor and successor is good enough to capture all
significantly correlated signals from two related streams. What this
means is that the predecessor must fire at least once for every ten
firings of the successor synapse.

But what happens if it fires more or less?

And I don't understand if you are talking about to inputs to a single
neuron, or an input to a neuron relative to when the neuron itself fires.

Is the point here to separate raw sensory input signals into multiple other
signals? How many? Do you have multiple layers of separation or is it
only one layer you are talking about. Is there an effective fan-out in
this process so that N inputs go to X*N separated outputs?

It's just not clear what you are talking about here.

2. Once signals are separated, the next task is to fuse them. This is
done by using special neurons that detect concurrency. Of course,
since the sensory space is uncertain, these neurons will fire even if
a fraction of its inputs did not fire concurrently.

And again, how are the signals selected for fusion and how are the
fusion "neurons" tuned? What controls the setting of the fraction?

Signal fusion is the simplest thing of them all. It is basically a
Hebbian process. Synapses (these may be chosen at random) that fire
together are strengthened while the others are weakened and eventually
severed. I use a 20 to 1 correlation ration for these neurons for
reasons that I will not go into. In my network, if 90% of the inputs
fire simultaneously, the neuron fires.

90% is a 10 to 1 ratio. What do you mean by "using a 20 to 1 correlation
ratio"?

As I said, it's not rocket science. And surprise, it does not even
require any advanced math. The hard part was to figure it out from
scratch. The good news is that it works like a charm.

I still don't really follow what you are doing here. Is the basic model
you are using like neurons with with multiple inputs? How many inputs do
you tend to use? A fixed number or variable? And do the separation
networks and signal fusion networks have multiple levels or single levels?

The concept of fusing signals that have greater correlations seems like it
makes some sense. But the concept of separation based on correlation
doesn't make much sense to me in terms of what you might actually be doing.

Do you used a fixed network for either of these, are does your system
dynamically add neurons, and remove them over time? Do you have some fixed
number of neurons in each network or is it somewhat arbitrary? How do you
know when the signals have been separated enough and don't know more
separation? And how do you know when to stop fusing signals?

In concept, I'm trying to do something very similar to what I think you
might be doing in your network, so I'm trying to understand just what you
have done.

It seems to me that this should require more than one separation followed
by one fusing. I say that because I believe for real world problems, I
don't believe it would be practical to separate all the signals down to all
their components. It might require a 1 to 100,000 fan out to do it. So
you start with a million inputs, fan out to 10^12 separations, and fuse
back to a million signals.

To prevent the network from having to support the 10^12 middle terms, you
should be able to separate and fuse through many different levels.

My pulse sorting network did that, but I don't think it was using the
correct separation and fusing techniques. It separated based on one
measure of correlation, but it fused by force (caused by fixed network
topology), not by some measure of correlation to determine whether the
signals were "worthy" of being fused.

--
Curt Welch http://CurtWelch.Com/
curt@xxxxxxxx http://NewsReader.Com/
.

Relevant Pages

  • Re: Symbolic AI: Why Marvin Minsky and Curt Welch Are Out to Lunch
    ... One can start with sensory processing. ... follows that sensory cortex neurons are signal separation processors. ... They separate specific signals from a given stream and sends them down ... fire about 10 ms before the target neuron fires. ...
    (comp.ai.philosophy)
  • Re: Symbolic AI: Why Marvin Minsky and Curt Welch Are Out to Lunch
    ... One can start with sensory processing. ... follows that sensory cortex neurons are signal separation processors. ... They separate specific signals from a given stream and sends them down ... fire about 10 ms before the target neuron fires. ...
    (comp.ai.philosophy)
  • Re: Symbolic AI: Why Marvin Minsky and Curt Welch Are Out to Lunch
    ... follows that sensory cortex neurons are signal separation processors. ... They separate specific signals from a given stream and sends them down ... fire about 10 ms before the target neuron fires. ...
    (comp.ai.philosophy)
  • Re: Neurons dont block so brain not a parallel computer?
    ... First off, feedback has to do with "loops", and you've not remotely ... neurons removed laterally can mutually ... of the signals", and effects depending on the neuron's state, is ... former is excitatory and the latter is inhibitory. ...
    (comp.ai.philosophy)
  • Re: Why Software Is Bad and What We Can Do to Fix It
    ... No one claimed that 100% reliability is possible, ... neurons to contain it all. ... relative temporal order of the arrival of the afferent signals. ... >> occurring in the brain at any one time, i.e., synchronously. ...
    (comp.object)