In a recent discussion ("Re: The 4 modes, active control
systems") Bill makes a very important remark:
[Bill Powers (970921.0728 MDT)]
[...] As I said to Marc Abrams yesterday,
the nature of neural signals says they can't go negative, so if you have
an excitatory reference signal and an inhibitory effect of the perceptual
signal at the comparator, a zero reference signal guarantees zero error
for all magnitudes of perceptual signal. The opposite arrangement (which
seems to exist in the brainstem) doesn't work this way. An inhibitory
reference signal, when set to zero, doesn't keep an excitatory perceptual
signal from generating an error signal. So "inverted comparators" can't
be turned off except through actively suppressing them with a large
inhibitory reference signal, larger than any perceptual signal that can
occur. The phenomenon of decerebrate rigidity shows this effect. If you
chop a live cat's brain off at the midbrain, all the reference signals in
the brainstem presumably go to zero -- and the result is full and active
extension of all the limbs. Grisly games these guys play, but the
information should't be wasted once you've wasted a cat to get it.
Here he mentions two variants of mismatch detection,
while he, unfortunately, overlooks the combination of the two:
OUTPUT
A
>
+->(EXC)<-+ REFERENCE 1 0 0 1
> A | INPUT 0 0 1 1
> > > =================
+->(INH)<-+ OUTPUT 1 0 1 0
> > MISMATCH 1 0 1 0
> >
INPUT REFERENCE
NB: the INHibitory neuron requires two inputs to be activated
the EXCitatory neuron requires one input to be activated
Without any knowledge of the details of PCT one can identify
a great advantage of the variant in which input only inhibits:
OUTPUT
A
>
(EXC)<-+ REFERENCE 1 0 0 1
A | INPUT 0 0 1 1
> > =================
+->(INH) | OUTPUT 1 0 0 0
> > MISMATCH 1 0 1 0
> >
INPUT REFERENCE
NB: If these pictures do not show right, check
out: http://www.xs4all.nl/~mervyn/bg.html
Since this gate can detect only half of the mismatches between
input and reference, a system that uses the output of such gates
will more easily remain focused on what it is trying to 'test' (its
reference). The output of the gate will provide less noise (due to
incompleteness of the internal model), while providing energy
for action in case of references are not 'mirrored' by the input.
It provides (limited, but correct) information on correctness, while
leaving out the noise that is caused by its incompleteness.
On the other hand, this gate demands that the reference and output
are not part of a single loop (a demand which is met by PCT). This
demand is due to the fact that if the system would be a closed loop,
the system would run out of energy (both to act and to 'refer') if
input would be maximal or the reference zero.
Despite the fact that I completely recognize the power of the PCT concept,
I would like to point out that one of its fundaments, the notion that an
*initial body of reference* is absolutely required to get a PCT-like
network going is not all that fundamental.
Actually, the opposite is true for a network that uses 'full' mismatch
detecting gates. If one gives up the advantage of 'focused' learning, one
receives the opportunity to: (1) start from scrap and (2) let the
reference be a transformation of the mismatch detector.
One can start from scrap because a system that is nothing can't control
anything. Therefore it has to learn to create references first. An infant
is exposed to a lot of input, but it has little physical control. Later
on it will preserve its (hopefully) healthy reference (a family, good
food, etcetera) by communicating and acting physically.
I personnally find a system in which the mismatch patterns are transformed
to become references more appealing than a system in which the reference
comes from a second system (or a hierarchy of systems) that is
energeticaly isolated.
In practice (cf. http://www.xs4all.nl/~mervyn ), such a system exposes a
different set of properties. Because of the explosive combination of full
mismatch detection and full feedback (mismatch -> reference -> ... ) this
tends to become relatively input independent (just like the reference in
PCT), while not giving up the possibility to be energized and informed by
the actual input (which poors in at zero-reference) or by the mismatch
pattern. It is the formation of this activity 'cloud' that I find
especially appealing, but that's just an opinion.
Regards, Mervyn