Alt-BEAM Archive

Message #00712



To: Noam Rudnick rudnick1@cwix.com, beam beam@corp.sgi.com
From: Sean Rigter rigter@cafe.net
Date: Sat, 20 Feb 1999 20:03:14 -0800
Subject: RE: associative memories


wilf wrote:

Noam,

We all think it's brilliant: opening up a door so to speak. Have you
tried
adding caps (like Nu neurons) to the input nodes? In theory that won't
change the DC behaviour but will slow down the change of state
transitions
so you can observe them. Use and RC of 1 about second ie 10M resistors
(effectively 2 in parallel) and 0.22 caps. I have attached a GIF
made from the "sketches" you send me.

wilf

PS you want me to cc my comments to the list?


Noam Rudnick wrote:

I'm glad you liked it =). I hadn't thought about the caps, i'll try
that too.
Thanks a lot for the circuit diagrams (when you say "sketches" I
definitely
take that as a compliment. I wasn't sure if you would even be
able to decipher my scribbles). In the middle sketch, is the
star-looking
thing... a transistor? Lastly, may i ask what the switch is for? Or is
that
just a way to switch between the two memory sets?
I'm glad that you liked it, and that my hours in the basement were not
wasted =)

It would be great if you could cc your comments,to the list because
I'm afraid that without any diagrams, my original ramblings were
rather
difficult to understand.
What do you think about adding diodes so the output of one neuron
cannot affect its input directly? Would this be benificial? Would it
really
change anything in the net's behavior?
I think it will be a while before this kind of memory actually gets
put into any robots, since it is a little on the complex side. It
could be
used in conjunction with a simple vision system to recognize objects
that it
sees frequently. The problem is how to adjust the resistors properly
so
the robot can really make its own memories as opposed to just random
values and the old "let's see what happens" method. Any ideas on how
this could
be implemented? maybe we would just hardwire it into the robot (A
little to unnatural for me. Imagine being born, seeign a picture of an
animal,
and automatically knowing it is an elephant).
Well, enough of my ramblings.

Thanks a lot,
Noam



wilf wrote:

The star thing is a 4066 bilateral switch used in reversers etc. You get
4 on a chip. It is a 2 way device that acts like a "dry" switch contact.
It sort of takes the place of the inverter with transistor. It has no
forward voltage drop at low current. The on resistance is about 10 ohm
and off is infinity. You can use it like a transistor (12V max) if the
current is not to high It is perfect for switching analog signals.
The switch in the second diagram just takes the place of the 4066 (or
your transistor). Try using a 10M resistor in series with the switch
with both in parallel to the existing 10M resistor. That way you
unbalance the node past the switching threshold but don't
charge/discharge the caps instantly. You would probably need 2 or 3 to
"stimulate" all useful node combinations in your exploration. When you
get interesting patterns or unpredictable behaviour you're on to
something. The more it "misbehaves" the more potential for BEAM
applications.

What was that hopfield doc URL again?

The image app is great (if it works) The input of one of n images
selects one of n instinctive BEAM mech responses. PUSSYCAT=DANGER:RUN
FOR IT! NOAM=MY CREATOR:I FLASH ALL MY LIGHTS FROM HAPPINESS:AND HOW MAY
I SERVE THEE (needs imprinting mechanism)?

Meantime I'll wire it up and try it for myself. I'll think about the
diode idea

wilf



Noam Rudnick wrote:

yeah, i wish i could get my robots to reckognize me =)

thanks,
Noam


wilf wrote:

Hi Noam,
thanks for the url
I tried your circuit but it oscillates like crazy because the inverters
are forced into the linear region. Maybe it's my layout or power
supply.
Do you use LEDs for indication or a scope? Anyway I'll play around some
more. A 74HC14 can be used but has thresholds ~ 1/3 and 2/3 Vcc so
would need 4 resistors per input. That means a minimum network size of
5
inverters each with 4 resistors at the output connected to the inputs
of
the other inverters. What you get is majority gate (voting) logic (I
call them nutrons). If use minimum 10 HC14 inverters with 9 resistors
you should be able to run LIFE on such a network. You know something
about CA?

later
wilf

------------------------------------------------------------------------
eGroup home: http://www.eGroups.com/list/alt-beam
Free Web-based e-mail groups by eGroups.com

Home