Multivariate local memory

[From Rick Marken (2015.04.19.1840)]

Control Hierarchy1.xls (27.5 KB)

···

Rupert Young (2015.04.19 17.00)

RY: In a basic control system its reference goal varies according to the value defined from above, so the same system services multivariate goals. With the addition of local memory (of perceptual signal which can be used later as a reference) what is actually being recorded? Is it a single value of past perception? In which case how can this system be used for anything other than that perceptual value, when memory is employed?

RM: This sounds like a great research question. The memory aspect of the PCT model is not very well specified. But I think your questions are answerable in the context of the existing model. (if I understand your question correctly). Note, in Figure 15.3 of B:CP, that the “memory switch” is between the output of the memory box (which presumably contains the addresses of the reference signal to be sent to all lower level systems) and the reference signal itself. The Figure shows the reference signal from higher level systems being sent to the comparator of only one lower level system. But that reference signal could be sent to the comparators of many lower level systems. So the same higher level system can service multiple (I presume that’s what you mean by multivariate) lower level system goals (references).

RM: Similarly, the lower level system in Figure 15.3 is shown as receiving a reference signal from only the one higher level system but, in fact, PCT assumes that the reference input to each lower level system is the sum or references from several higher level systems (see my spreadsheet hierarchy model: which I’ve attached). So the complete"picture" in Figure 15.3 would show multiple reference signals entering the comparator coming from multiple higher level systems.

RM: The reference signal from the higher level system shown in Figure 15.3 actually enters the comparator of the lower level system when the “memory switch” is not thrown: that is, when the memory switch directly connects the output of the the memory box to the reference signal. However, when the memory switch is thrown (as pictured in Figure 15.3), the perception that is “played back” into the perceptual path of the higher level system is just the perception that corresponds to the single output of the memory box. That reference corresponds to the single perception that the higher level system requested.

RM: All that is theory, of course. What we need now is to create an experimental situation that corresponds to at least a two level hierarchy with at least two control systems at each level. I’ll try to come up with a control task that would require such a model. Then we would have to think of a way of testing the memory model in 15.3. That model predicts that the subject should be able to recall only the single perception sent to the lower level system and not the combination of references sent to that system. I’ll let you know if I come up with anything. This is just what I hoped CSGNet could be used for; to develop experimental tests of the PCT model, particularly those aspects of the model that have received little or no tests so far.

Best

Rick

For example, suppose you are looking for different stars in the night sky. You remember their respective locations (bearing), with respect to some initial point. You are able to change your gaze to fixate on each one in turn, from betelgeuse, to polaris, to sirius etc. Maybe there are a couple of levels, at least, here, one for the stars and one for the gaze direction. How is this mulltivariate knowledge incorporated within the PCT memory model? Does the star control system “encode” memory of all stars? Or is there a different control system for each star? If the direction control system memorises a single bearing perception how can it control to a different direction, when based on memory? Does the direction system need memory?

A few queries which don’t seem to be answered by the model, as I understand it. Any thoughts?

Regards,

Rupert


Richard S. Marken

www.mindreadings.com
Author of Doing Research on Purpose.
Now available from Amazon or Barnes & Noble

[From Rupert Young (2015.04.24 21.00)]

(From Rick Marken (2015.04.19.1840)]

Actually what I mean is, is just a single value (of the perception)

stored (to be retrieved later) for the local memory or multiple
values? If the latter then how is this represented in a neural
(control) system? And what is the mechanism for address retrieval?
This would seem to imply that there would be unlimited storage.

Any thoughts on how it would work with my simple stars example?

So, a memory is activated from a higher system which can then flow

back up the hierarchy to form part of a perception of another higher
system?

Ok, that'll be interesting, and would be keen to collaborate.

What I'd like to get to is ways of implementing a memory model in

robots. I can think of some software ways of doing it but would like
to understand the biologically-plausible, PCT way. It would be
interesting to build an associative memory model in PCT.

Regards,

Rupert
···
          Rupert

Young (2015.04.19 17.00)

          RY: In a

basic control system its reference goal varies according
to the value defined from above, so the same system
services multivariate goals. With the addition of local
memory (of perceptual signal which can be used later as a
reference) what is actually being recorded? Is it a single
value of past perception? In which case how can this
system be used for anything other than that perceptual
value, when memory is employed?

          RM: This sounds like a great research question. The

memory aspect of the PCT model is not very well specified.
But I think your questions are answerable in the context
of the existing model. (if I understand your question
correctly). Note, in Figure 15.3 of B:CP, that the
“memory switch” is between the output of the memory box
(which presumably contains the addresses of the reference
signal to be sent to all lower level systems) and the
reference signal itself. The Figure shows the reference
signal from higher level systems being sent to the
comparator of only one lower level system. But that
reference signal could be sent to the comparators of many
lower level systems. So the same higher level system can
service multiple (I presume that’s what you mean by
multivariate) lower level system goals (references).

          RM: Similarly, the lower level system in Figure 15.3 is

shown as receiving a reference signal from only the one
higher level system but, in fact, PCT assumes that the
reference input to each lower level system is the sum or
references from several higher level systems (see my
spreadsheet hierarchy model: which I’ve attached). So the
complete"picture" in Figure 15.3 would show multiple
reference signals entering the comparator coming from
multiple higher level systems.

          RM: The reference signal from the higher level system

shown in Figure 15.3 actually enters the comparator of the
lower level system when the “memory switch” is not thrown:
that is, when the memory switch directly connects the
output of the the memory box to the reference signal.
However, when the memory switch is thrown (as pictured in
Figure 15.3), the perception that is “played back” into
the perceptual path of the higher level system is just the
perception that corresponds to the single output of the
memory box. That reference corresponds to the single
perception that the higher level system requested.

          RM: All that is theory, of course. What we need now is

to create an experimental situation that corresponds to at
least a two level hierarchy with at least two control
systems at each level. I’ll try to come up with a control
task that would require such a model. Then we would have
to think of a way of testing the memory model in 15.3.
That model predicts that the subject should be able to
recall only the single perception sent to the lower level
system and not the combination of references sent to that
system. I’ll let you know if I come up with anything. This
is just what I hoped CSGNet could be used for; to develop
experimental tests of the PCT model, particularly those
aspects of the model that have received little or no tests
so far.