In the HPCT model can upper levels
set the gain as well as the reference
level for lower levels?
[From Bill Powers (2003.01.06.0826 MST)]
Bruce Gregory (2003.01.06.1003)–
Yes. I’ve mentioned this possibility a number of times on the net –
higher levels adjusting parameters (gain is one parameter) of lower
levels while monitoring aspects of control performance rather than the
states of ordinary controlled variables. Tom Bourbon demonstrated
(something like 20 years ago) a simulation of a higher system that
monitored the error signal in a control system and reorganized the output
gain of the target control system to achieve minimum error
(“reorganized” meaning it randomly varied the gain up and down
in the manner of e. coli). A systematic method would work, also. Most
people have enough trouble understanding control via reference signals,
of course, so I haven’t spent any significant time on control via
parameter variations. I’m toying with the idea of introducing it in a
model of multidimensional control I’m working on. It’s the PCT
alternative to the Extended Kalman Filter and all that Modern Control
Theory jazz.
I have been
reading (and recommend very highly)
Cortex and Mind: Unifying Cognition by Joaquin Fuster. I’ve been
impressed by the ability of neural systems to inhibit the firing of
other systems at apparently the same level in the hierarchy. (This
occurs thoughout the neural system, but is clearly demonstrated in
the
visual cortex in the determination of edges, for example.)
“Inhibition” is a catch-all term that covers several possible
effects of a neural signal on a neuron. In the spinal motor neuron, it
refers to the subtraction of a negative feedback signal
(originating in a Golgi receptor) from a positive input signal, the alpha
motor signal from higher systems. This makes the motor cell into a
comparator, as well as an output amplifier.
Inhibition can also mean division, in the sense that the magnitude
of the inhibiting signal alters the slope of the relationship between an
input signal and an output signal – the relationship we call
gain. The greater the inhibiting signal, the lower the slope or
gain. “Facilitating” signals have the opposite effect on gain.
When I speak of signal magnitudes, I always mean frequency of firing, by
the way. not single-impulse measures.
Still another meaning is gating. In this case, the inhibiting
signal turns a neuron completely off, so inputs produce no output at all.
The implication of gating is that the neuron works like a binary device:
either the input produces output, or it doesn’t. Gain adjustment,
sometimes called modulation, is the most general category for
multiplicative effects, with gating being an extreme on a continuum of
effects. You can get an effect like on-off switching if there is positive
feedback from output back to input – see chapter 2 of B:CP for more
possibilities.
Subtraction and gain reduction are the two main meanings of inhibition,
as far as I’m concerned. Both functions can be useful anywhere in a
control system, from its input function through its comparator to its
output function. I know that neurologists have made much of the special
significance of inhibition (there’s even a big thick book titled
Inhibition, which to me is like someone publishing a big thick
book called Addition). But I think it’s just one aspect of the
analog computing that goes on everywhere in the brain, and nothing to get
very excited about for its own sake.
Applying this
model to PCT suggests to me something like the following. When I
catch
something in my peripheral vision, I turn my head. Since I can not
do
this and control my perception of distance to the car in front of
me,
the gain on this latter control loop in lowered (inhibition) and not
restored until I again look forward.
I think that’s a good possibility. One has to be careful about saying
exactly what is inhibited and where in the control loop the inhibition
takes place. If you inhibit the perceptual function, for example, the
perceptual signal becomes smaller, and this would increase the error
signal and cause an increase in output activity, which isn’t what
you want. If you inhibit the output function, the output action (the
reference signals reaching lower systems) will go to zero, and you have
to ask seriously if this will do what you want done, either. It may be
that what has to be inhibited in cases like this (and I agree that
inhibition is probably the right process to consider here) is something
like the input to an integrator, where a zero input means that the
integrator will hold its output constant or nearly so for a while, until
it receives nonzero positive signals to make it larger or negative
signals to make it smaller. This means that whatever action is going on
neither increases nor decreases while the (total) inhibition is in effect
(though there might be some small drift, because neural integrators won’t
be perfect, and this would lead to some testable predictions).
Where I suspect a
hierarchy is clearly at work is in my decision not
to look left until the way in front of me is clear. Again this
appears
to involve gain rather than reference level.
What do you think?
I think these are good suggestions. The decision to look or not look,
being binary, makes me think of logical processes, but of course those
have to be translated into specific states of analog signals at lower
levels to have any useful effect. A change of gain in an analog system
(in the right place) might well result from the logical conclusion that
you can’t look to one side and straight ahead at the same time.
Best,
Bill P.