Re: Temporal Learning

"JGCASEY" <jgkjcasey@xxxxxxxxxxxx> wrote:
> You have gone all silent on me since I offered
> you a challenge Curt ;)

I started on creating a simple hard-coded answer,
I've just been too busy to finish it. Here's what I
got so far:

H E L L O
. . . . . . - . . . - . . - - -

(A) |-|-|-|---|---|-||-|-|---|-||-|-|----||-||-||

1.5L of (A)
(B) | | | | | | | | | | | | | | | | (one pulse per sym)

1.5H of (A)
(C) | | | | | (one pulse per dash)


3.5L of (B)
(D) | | | | | (one pulse per letter)


> Essentially your net is a gap sorter.
>
> IF gap > x Then Word_Gap
> IF gap > y Then Letter_Gap
> IF gap > z Then Pulse_Gap
>
> o is a node and X the input node.
>
> X-----o-----o----> word gap
> | | |
> o-----o-----o----> letter gap
> | | |
> o-----o-----o----> dot/dash gap
> | | |
>
> Separating the dash duration from the dot duration
> is the next biggy.

Yeah, above you see I created a signal for the dash. It's easier to do
this if you have a very simple and formal (aka exact) signal encoding. In
real life, there has to be more flexibilty which requires a lot more
samples (than two per dash), but the net should be able to do it.

> A constant stream of pulses with
> gaps within whatever range you choose will cause
> the outputs to be swept until it settles.

RIght, if you don't have other inputs, the net goes "dead" between pulses
and can't create any type of outputs. A seperate input with a constant
stream of pulses could be used to "power" the logic in the net between
pulses. Or, you could just add background noise to the single input
stream.

This style net is really designed in my mind to work with a lot of data, so
that something like the idea of a "dash" is not represented by a single
pulse, but instead, but lots of them. But that makes it much harder to
understand what it is doing when trying to hand code an example.

However, with the help of the extra constant data stream, combined with the
input data, you can create NOT effects which can lead you to separate dot
and dash signals I think. And also allow you to create an "end of letter"
signal. Running those though a series of nodes with gap lengths near the
dot+dot vs dot+dash vs dash+dash combined lengths I belive would allow the
system to create a sieve. Which when the "end of letter" signal passes
through it later, would come out different place for each combination of
dots and dashes seen in the letter.

But, I've only tried to do this typeing on the computer and it's hard to
figure it out that way. I need to sit down with a pen and papper and
figure it out than then post the answer.

> Thus each
> output could be a duration measure. The longer it
> sweeps the longer the constant frequency has been
> applied. A rather ad hoc and convoluted solution,
> if it is a solution. The next step is using the
> outputs as inputs to what? Another net or further
> down the net "somehow" a set of node values to
> divert the pulses to one of the 26 possible outputs
> desired?

Yeah. Somehow. :) The problem is that the net, when learning for itself,
has the power to "find" answers that are not at all obvious to us at first.
So, what I manage to come up with by hand, might not in fact be anything
like what the net would naturally create on it's own with training.

And since I haven't spent anywhere near the time I need to spend struyding
what type of solutions it does come up with with similar problems, I'm not
very good and knowing what it might do yet.

However, this type of problem is simple enough that it should be reasonable
to just do what we are doing here and find a hand coded answer just to
prove that some answer is possible for this type of problem and ignore the
issue of whether it can create an answer on it's own for later.

I'll see if I can find more time to work on this....

> Regards,
>
> John Casey

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

Relevant Pages

  • Re: Mechanical Dualism versus Naturalized Epistemology
    ... >>> desired output for a given input pattern of pulses. ... > values and the gap values (which depend on the input ... But it gets complex because as you sort ... And as signals merge, new gaps get defined so it's sorting pulses ...
    (comp.ai.philosophy)
  • Re: What is missing from AI ?
    ... All cross connects are ... One node will force all pulses < 20 to the same place. ... Now you have a toploty that can easily cross signals and route any hidden ... them are as important to me as the reward prediction probelm. ...
    (comp.ai.philosophy)
  • Re: Does artificial intelligence require pleasure?
    ... if (time>= gap) { ... split the set of all pulses in half. ... reaction network. ... and had it interacting with a real environment. ...
    (comp.ai.philosophy)
  • Re: Trying to Falsify a non-scientific theory?
    ... published your gap stuff. ... produce such signals. ... producing high-level biosystems since humans have in fact produced ... restrictive model for the manufacturer of the signals, ...
    (talk.origins)
  • Re: IR LED direction accuracy
    ... detector with some basic AGC and then a filter at 1Mhz. ... remote control receiver module. ... Minimum & maximum number of pulses in a burst as well as ... doesn't match the signals that are ambient. ...
    (sci.electronics.design)