What if?

I've been reflecting about the nature of HPCT, specifically, the hierarchy.

What if it's not a "hierarchy"? What if it's flat? What if it's better
represented perhaps by a set of concentric circles? Could I still have the
same set of dependencies in a flat, concentric-ring representation? Might
the output of inner-circle control systems have the same relationship to
"outer" as to "upper"?

Just wondering if there's a particular, unbendable reason for the vertical
nature of a hierarchy.

Regards,

Fred Nickols
"Assistance at a Distance"
nickols@att.net
www.nickols.us

[Martin Taylor 2006.08.17.15.33]

Fred Nickols (apparently Thu, 17 Aug 2006 15:14:36 -0400) said:

I've been reflecting about the nature of HPCT, specifically, the hierarchy.

What if it's not a "hierarchy"? What if it's flat? What if it's better
represented perhaps by a set of concentric circles? Could I still have the
same set of dependencies in a flat, concentric-ring representation? Might
the output of inner-circle control systems have the same relationship to
"outer" as to "upper"?

Just wondering if there's a particular, unbendable reason for the vertical
nature of a hierarchy.

The layout of a diagram is just a graphical convenience for easy understanding. A hierarchy can be drawn in lots of ways, most of which are harder to understand that the "vertical" arrangement.

What do you imagine the connections to be between one circle and next one inside or outside it? Assuming the outer world to be outermost, if the connections are such that perceptual outputs go to perceptual inputs of the next circle inwards, and outputs go to the reference inputs of the next circle outwards, then the circles are the same as the hierarchy.

What would it mean to be "flat"? Does that mean all the elementary control units are connected _directly_ to the sensors and muscles, and non connect to the perceptual outputs and reference inputs of other control units? If that's the implication of "flat", then I think there are very good reasons why this is a highly improbable organization. Is the perception of "freedom" as tightly linked to the current state of your sensors as is the perception of "bright"?

Its quite another matter as to whether the pure hierarchy (laid out vertically or in concentric circles) represents realistic living systems. In my own mind, I separate PCT (which I think has to be valid) from HPCT (which needs to be tested much more than just to show that high-level control systems can control through setting reference levels for lower-level systems). HPCT is plausible, whereas PCT comes close to being a provable consequence of basic physics principles (as I attempted to show in my Editorial for the PCT issue of the International Journal of Human Computer Studies).

What HPCT asserts that might or might not be true is that there is a strict hierarchy -- meaning that no perceptual function of an elementary control unit at level N sends its output signal directly or indirectly to the input of any perceptual function at level N or below (other than by way of its own or other output functions), and that no output function of an elementary control unit at level N sends its signals to the reference input of any unit at level N or above (other than by way of soem perceptual input function). Bill has argued for an even stricter criterion: that if output connections were to sometimes skip a level (level N outputs going to references at level N-2, for example), the level-skipping connections would less effectively replicate the functions of connections that went through the intermediate levels, and thus would dissapear through evolution (I hope this correctly states and doesn't oversimplify Bill's thought).

The HPCT assumptions need not be true, but if they are not, testable consequences should follow. Peter Small and others have argued for connections from perceptual function outputs directly to perceptual function inputs at the same level, creating dynamic loops that could have interesting attractor dynamics. I have argued for the possible existence of a parallel structure that handles functions that HPCT lists as being at and above the category level; the connections of this structure go to and from all levels of an HPCT-like hierarchy. Again, there ought to be testable consequences if that structure is more nearly correct than HPCT. (I haven't thought of such consequences, so apart from mentining the possibility a few years ago, I haven't tried to push the concept.) In the kind of language you used in the quote above, I'd think of it as vertical, but with side flaps!

My somewhat discursive point is that the "vertical" layout of the hierarchy means nothing, but the conenction structure of the hierarchy is a hypothesis that potentially could be tested against sufficiently well described alternative hypotheses.

Martin

[From Bill Powers (2006.08.17.1655 MDT)]

I've been reflecting about the nature of HPCT, specifically, the hierarchy.

What if it's not a "hierarchy"? What if it's flat? What if it's better
represented perhaps by a set of concentric circles? Could I still have the
same set of dependencies in a flat, concentric-ring representation? Might
the output of inner-circle control systems have the same relationship to
"outer" as to "upper"?

The idea of the hierarchy came about pretty early, from considering the spinal reflexes, which are organized as control loops. Those loops are built in; there's no way for the brain to operate the muscles except through them. When you trace out the circuits (as shown in B:CP), it's clear that the spinal motor cells are comparators, and that the so-called command signals reaching the motor cells from higher up are reference signals. I don't think there can be any doubt about that, and neither did people like Houk in the 1950s.

It slowly grew on me that the whole central nervous system could be organized that way. At the level of spinal reflexes there are at least 800 control systems, and probably quite a few more than that because different branches of a muscle can be separately energized. This meant that you could trace input from that level and outputs to that level, so the spinal systems became just like an environment of the rest of the brain, and you could then start over to define a second level of control, and so on for however many levels there really are. From then on the job became one of trying to see if there was any observable evidence for such levels, and what sorts of control systems might exist at each level. That was a slow process! Five levels in 1960, nine in 1973, eleven sometime in the 1980s, and none added since then.

There are circumstantial reasons for believing this organization. One is developmental. If the brain's organization matures from the bottom up, each new level could become organized and operate by itself, either with inherited values of reference signals or simply absent reference signals, equivalent to a value of zero. Each new level then becomes the world which new levels control. There is never any need for a higher system to supercede, replace, override, or "subsume" lower systems (the Brooks architecture). New levels can be added without requiring changes in the existing levels (though such changes would undoubtedly occur). The point is that adding levels doesn't call for total reorganization of everything below them.

Another reason is evolutionary. At earlier stages during evolution, it stands to reason that each level we now find in a human brain, or think we find, was the highest existing level. So we should be able to find organisms with only the intensity level, or with only the intensity and sensation levels, and so on. I never got past the stage of idle speculation about that, but it's interesting to look at other organisms with this in mind. The hierarchy organization makes it possible for viable organisms to exist with any number of levels from one upward.

Finally, neurophysiologists long ago realized that there is more than one level of reflexes; there are spinal reflexes, brainstem reflexes, and midbrain reflexes -- and possibly still more. Each level necessarily works through the systems that make up the lower reflexes. Not much circuit analysis has taken place; mostly neurologists chase impulses around, whereas what we need is for neural functions to be analyzed -- how one set of signals consisting of trains of impulses get turned into new signals. I see studies of that kind only rarely in Science and Nature, but even then nothing like what we need.

Martin Taylor has taken care of the geometric aspects of your question. As he says, it's not the geometry that matters, but the connectivity.

Besgt,

Bill P.