Alt-BEAM Archive
Message #00642
To: BEAM beam@corp.sgi.com
From: Noam Rudnick rudnick1@cwix.com
Date: Fri, 19 Feb 1999 11:26:43 -0500
Subject: [alt-beam] associative memories
content-transfer-encoding: quoted-printable
content-type: text/plain;
I have been experimenting with memory in beam, specifically, associativ=
e memories. These can be built just like inverter flip flops in a sense, e=
xcept with more inverters. Two "memories" are stored in a standard flip fl=
op (0utput of 1 inverter connected to input of the other, through a resisto=
r, and the other way around.). When a low, for intance, is applied to the =
input of one inverter, then the network can remember that the other node sh=
ould be high. This basic concept can be applied to X number of nodes as lo=
ng as the input of every node is connected to the output of all the other n=
odes/neurons, but not connected to itself. The resistor values can be tota=
lly random, or hand-picked; it doesn't matter.
I tried this "associative neral net" with a total of three inverte=
rs, and got the expected results. There were two "stable states" the netwo=
rk could have at any time: 010, or 101. Then, just for the hell of it, I t=
ried to see how easy it was to have these memories change. i built a simpl=
e circuit somewhat similar to a PNC. Basically, it consisted of a large ca=
pacitor that had one connected to ground, and the other connected to the in=
put of an inverter. Also connected to the input of the inverter was a larg=
e resistor connected to +5v. The output of the inverter connected to the b=
ase of a transistor, which replaced one of the resistors going from the out=
put of one neuron, to the input of another.
So then I powered the sucker up. At first, I had the normal 010, and 1=
01 stable states, but after a while, the network changed. Now there were t=
wo totally different stable states (I can't remember them off the top of my=
t head, but they were different from the first two). Anyway, whenever i wa=
nted to, I could aply a positive voltage to the input of my "pnc" neuron, a=
nd it would return to having the original stable states...for about a minut=
e or so, after which the network would "forget" them.
-So what do you all think?
-Do associative neural nets have a place in beam?
-Will our bots finally begin to be equiped with memory?
-I'd be glad to hear your thoughts on this stuff.
Thanks,
Noam Rudnick
------------------------------------------------------------------------ZZX=
X
eGroup home: http://www.eGroups.com/list/alt-beam
Free Web-based e-mail groups by eGroups.com
content-transfer-encoding: quoted-printable
content-type: text/html;
the =
expected results. There were two "stable states" the network cou=
ld =
have at any time: 010, or 101. Then, just for the hell of it, I tried to s=
ee =
how easy it was to have these memories change. i built a simple circuit =
somewhat similar to a PNC. Basically, it consisted of a large capacitor th=
at =
had one connected to ground, and the other connected to the input of an =
inverter. Also connected to the input of the inverter was a large resistor=
=
connected to +5v. The output of the inverter connected to the base of a =
transistor, which replaced one of the resistors going from the output of on=
e =
neuron, to the input of another.
So then I powered the s=
ucker up. =
At first, I had the normal 010, and 101 stable states, but after a while, =
the =
network changed. Now there were two totally different stable states (I can=
't =
remember them off the top of myt head, but they were different from the fir=
st =
two). Anyway, whenever i wanted to, I could aply a positive voltage to the=
=
input of my "pnc" neuron, and it would return to having the origi=
nal =
stable states...for about a minute or so, after which the network would =
"forget" them.