Re: Is the Curt net a kind of decision tree?
- From: curt@xxxxxxxx (Curt Welch)
- Date: 30 Jun 2006 22:14:58 GMT
"JGCASEY" <jgkjcasey@xxxxxxxxxxxx> wrote:
"JGCASEY" <jgkjca...@xxxxxxxxxxxx> wrote:
Curt Welch wrote:< ... >
This lack of power is the main problem I ran into when I was
working on your suggestion of trying to hand-wire a Morse code
detector. The very simple problem of trying to tell Dot Dash
from a Dash Dot was something I couldn't figure out how to do
with only a small number of nodes. I could make it produce
different behaviors (outputs) for Dot, and for Dash, but I saw
no easy way to make it produce the "A" output for the Dot Dash
sequence vs the "N" output for the Dash Dot sequence.
I suspected that it could be done but would require the help
of external state storage systems with feedback. But that I
think is overkill. The basic learning network needs to be able
to recognize a dot dash from a dash dot without external help -
the ability to recognize temporal patterns like that and respond
differently to various patterns must be fundamental to this type
of system.
I see this as the general segmentation problem although in this
case you are segmenting temporal events, as is the case in music,
speech and of course all the sounds of nature.
It has to deal with relative temporal data. For example let's say
we feed it dashes and dots,
__|||||__||__|||||__||________
will it then recognize this,
__||||||||||__|||||__||||||||||__|||||_____
as the same thing although the dashes are now the size of dots?
Then of course there are different frequencies that could be
modulated with the dash and dot data which should be recognized
as well without needing any more learning. It is essentially,
I guess, recognition of different durations of some stimulation.
You might modulate light instead of sound, or modulate the
pressing of someone's skin with dots and dashes.
Yeah these are all good points.
However, I don't think it's valid to say "with NO additional learning".
Many things like this what we can learn to recognize quickly do take some
time to learn.
I remember the first time I saw a 7 segment numeric display. It was hard
at first to see them as numbers. But it didn't take long to become
accustomed to the format and see them as numbers instead of odd square
shapes.
There does however seem to be an inherent pattern matching systems at work
that would naturally equate Morse code at different speeds as being the
same however.
This information is destroyed by your nodes when a pulse passes
through them.
Well, it's destroyed when two paths merge into one before being
sent to the pulse. That's where the damage is done. That and
the fact that the nodes just don't have any ability to make
decisions based on two separate pulse sources.
Of course "merging" is what it means to classify in that a set of
input patterns are "routed" the same way due to something they
have in common and the selective mechanisms that decide which way
to route things. My "frequency sorter" made use of the fact that
your nodes sort according to gaps values. So I could merge a range
of gap values to the same output.
The pulse that comes out of the node provides no
information as to where it came from.
That's true as well, but I don't think the down stream nodes
will need to know that. The pulse could in theory carry with
it a history of its entire path through the net up to that
point but I don't think that would be needed.
I don't see a pulse in your network being able to "carry with
it a history of its entire path through the net". Show me how
you can restore its history from a single pulse coming out of
a particular output?
I wasn't trying to imply it did it. I was just saying that was a possible
enhancement. A network which made decisions based on the entire path could
be built. I don't see much value in that however.
The best I could say is it came from a
limited range of inputs via a limited range of nodes, but which
ones exactly cannot be deduced from the output from whence the
pulse exits. Of course a feedback system connected to every node
could "recognize" the nodes involved by their count values.
< ... >
Now you need to work on how to get your net, or some version
of it, to discriminate between the order of two or more events.
Right.
Indirectly, it had powers to do it already. That's because there are
always other pulses to compare to. So even though it can't tell the
difference between AB and BA, it can tell that ABC and BAC are different
because of the different gaps between A and C and B and C in those cases.
However, it couldn't directly tell which was which and could only
indirectly do it by comparing the AC and BC gaps to a lot of constants. It
just seems to me that is to hard for something I think the net should be
better at.
Oh, and perhaps discriminate between durations some stimuli?
Well, the current net does nothing but discriminate durations. And it's
fairly easy to make it break stimulus signals down so it can do direct
comparisons between the start of each stimulus. But it can't compare the
duration of one stimulus to the duration of another in any direct
meaningful way which I think is just another example of it's lack of
ability to compare two separate signals vs two separate pulses on one
signal which is what the current design is based around.
--
Curt Welch http://CurtWelch.Com/
curt@xxxxxxxx http://NewsReader.Com/
.
- Follow-Ups:
- Re: Is the Curt net a kind of decision tree?
- From: JGCASEY
- Re: Is the Curt net a kind of decision tree?
- Next by Date: Combining AI techniques
- Next by thread: Re: Is the Curt net a kind of decision tree?
- Index(es):
Relevant Pages
|
