Re: Representing knowlegde about pattern recognition
- From: curt@xxxxxxxx (Curt Welch)
- Date: 17 Aug 2007 07:28:13 GMT
JP <gms2004@xxxxxxxxx> wrote:
On Aug 15, 7:55 pm, heiska mikko <mikk2...@xxxxxxxxxxx> wrote:
AI has to be able to learn and develop new ways of recognizing
patterns. AI must have modifiable templates for describing state
classes. It is not enough that computer code does different kinds of
recognizing tasks. AI must have those recognized pattern types
described with same kind of symbolic nodes and links as any other
knowledge / understanding-data. Then, computer code can be made from
that.
When learning to transcribe or understand voices of new language, AI
could test thousands of different recognition templates, guided by
some rough metrics indirectly indicating meaningfullness of text.
IMO pattern creation (not recognition) is just a tool to reduce the
amount of information in order to process it.
IOW if the amount of information perceived is "N" and the processing
capacity is "N-1", then pattern creation is the tool to reduce the
amount of information to the (a) lower level.
JP
I think that's exactly right.
The brain builds (and or tunes) pattern recognition devices as it's way of
creating patterns. And it does it in a way that helps to maximise the
amount of information it's able to represent with a finite number of
pattern recognition devices (or templates as heiska mikko said).
I believe in the case of the brain, the information flowing into it is not
just slightly greater than what it can represent, but that it's actually
many orders of magnitude greater than what it can represent in these
pattern recognition devices.
If you look at raw information flowing into the brain as pulse signals,
you can talk about how much information is represented in each signal by
translating it to something like bits per second. But the problem of
producing behavior when the sensory signals to not represent the entire
current state of the environment, is that the system must include memory
function which attempts to store as much of the information flowing in as
possible. This can be done by creating temporal pattern recognition
devices. The "temporal" part is key here because a temporal pattern
recognition device acts as a storage device.
That is, if you have an input sequence of tokens such as A B C, and you
have a temporal pattern recognition device looking for that pattern and
producing the output X when it sees the ABC pattern, then the current
output of the X is an indication of what as recently happened in the past.
It in effect includes a memory of the A, the B, and the C.
The goal of the system is no doubt to create (or tune) a finite number of
temporal pattern matching devices, to represent as much historic
information as possible in the sensory data.
By changing the devices together, you can represent more historic
information with less hardware. For example if you ABC was matched as M,
and DEF was matched as N, and M N was matched as X, and N M was matched as
Y, the X symbol has the historic meaning of ABCDEF. and the Y symbol has
the meaning of DEFABC. So with hardware using only 10 inputs, we produced
pattern meanings that would have required 12 inputs if done in a flat
network (two patterns of 6 inputs each).
But these examples make it look like a simple problem of sequencing which
it's not. With spike signals, it's not just sequence which is important,
the actual timing (temporal spacing) of the pulses are equally important.
So the whole trick to building the temporal pattern detection is to
compress the data and represent as much historic information about not only
the pulses, but their temporal timing as well.
So you might have one pattern detector to represent AAA happening close
together, and another pattern detector to represent A A A where the pulses
are spaced further apart.
I don't yet fully understand how to create, or tune, temporal pattern
recognition devices to maximize the amount of information represented with
a chained network of devices, but that's the problem I spend most my time
thinking about.
In addition to the default behavior of having them self-tune for the
purpose of maximising the amount of historic information represented, they
also all need to be trained by reinforcement. But it's the selection of
the patterns for the purpose of information maximising that's the most non
obvious to me and the problem I spend most my days thinking about.
One though that is materializing (again) as I write this, is the idea that
as you go higher in the pattern recognition hierarchy, the important
concept is not just that the physical receptive field will get larger (i.e.
the scope of raw sensory inputs that are part of the pattern), but that the
temporal receptive field needs to expand as well - meaning that the
temporal patterns the device responds to extends further back in time as
well.
--
Curt Welch http://CurtWelch.Com/
curt@xxxxxxxx http://NewsReader.Com/
.
- References:
- Representing knowlegde about pattern recognition
- From: heiska mikko
- Re: Representing knowlegde about pattern recognition
- From: JP
- Representing knowlegde about pattern recognition
- Prev by Date: Re: Is technology part of evolution?
- Next by Date: Re: Is technology part of evolution?
- Previous by thread: Re: Representing knowlegde about pattern recognition
- Next by thread: Re: Representing knowlegde about pattern recognition
- Index(es):
Relevant Pages
|
