B:CP Chapter 15: Memory & Address Signals

[From Chad Green (2014.08.21 15:03 ET)]

Erling, I wanted to let you know that I just shared your observations below with Francis Heylighen of the Global Brain Institute (http://pespmc1.vub.ac.be/HEYL.html )
since he had expressed a strong interest in Hawkins’s book On Intelligence.

Best,

Chad

···

Chad T. Green, PMP

Program Analyst
Loudoun County Public Schools
21000 Education Court
Ashburn, VA 20148
Voice: 571-252-1486
Fax: 571-252-1575
Web: http://cmsweb1.loudoun.k12.va.us/page/1966

“The future is already here—it’s just not very eevenly distributed.â€? - William Gibson

From: Control Systems Group Network (CSGnet) [mailto:CSGNET@LISTSERV.ILLINOIS.EDU]
On Behalf Of Erling Jorgensen
Sent: Saturday, November 16, 2013 2:19 PM
To: CSGNET@LISTSERV.ILLINOIS.EDU
Subject: B:CP Chapter 15: Memory & Address Signals

B:CP Chapter 15: Memory & Address Signals

[From Erling Jorgensen (2013.11.16 1100EST)]

I’m a little behind in keeping up with the chapters.

There is a piece of Bill Powers’ chapter on memory that is only slowly coming into resolution for me. It strikes me as a critical piece, but I haven’t seen much discussion of it. By way of forewarning, this turned into an extended essay,
touching first on Power’s understanding of reference signals, and then examining Jeff Hawkins’ treatment of the neurophysiology of the cerebral cortex of the brain.

I’m using the first edition of Behavior: The Control of Perception (1973).

On p. 217, Bill says, almost in passing:

“This, incidentally, solves the problem of translating an error in a higher-order variable into a specific value of a lower-order perception, a problem that has quietly been lurking in the background.�

I’d like to unpack what I think is going on in this passage and in the surrounding pages of B:CP.

First, what problem is being solved? And why is it a problem in the first place? To give it a name, it is the issue of commensurability. Wikipedia defines commensurability as follows: “Two concepts or things are commensurable if they
are measurable or comparable by a common standard.�

It’s the problem of interface. How will two things communicate in a way that is understandable to both sides? What language or units or frames do they hold in common?

At this juncture in the history of Perceptual Control Theory, this may seem a trivial problem. After all, we have already had numerous rigorous demonstrations of principle, in the various computer simulations that show just how well a
negative-feedback control architecture can operate. There are even simulations of hierarchical arrangements of control loops, where different classes of perception are being defined by the perceptual input functions, but stable control is achieved nonetheless
as higher levels set reference standards for lower level loops.

In all these simulations, commensurability between the different levels is presupposed. It’s handled by the computer. The equations assign numbers to the different variables, and the computations proceed straightaway. It is why we operate
with an assumption that variables perform as scalar quantities, even if they may represent a more complex form of perception.

That scalar assumption is itself based on a decision that Bill sets out on page 22 of B:CP:

“As the basic measure of nervous-system activity, therefore, I choose to use neural current, defined as the number of impulses passing through a cross section of all parallel redundant fibers in a given bundle per unit time.� (italics omitted)

Bill was an engineer, and he knew you had to pay attention to the units and definitions of the variables, so that they could be properly compared and manipulated. So again, the issue of commensurability seems to be solved, by how Bill
sets out his definitions. When we lay out equations for how control loops operate, so says he, we’re talking about “neural currents.� That is the common measure that will make the different layers commensurate.

This is a credible (and insightful) way to map the concepts of engineering control theory onto the biological mechanisms of organisms with their nervous systems. Whenever we build maps, we use what can be called “convenient fictions,�
which simplify but hopefully retain essential features of the territory we are trying to map.

The problem that Bill knew was “quietly lurking in the background� was that it can sometimes be hard to squeeze the underlying details into the convenient fictions we choose for our map. In the case of nervous system operation, there is
a sizeable risk of incommensurability, as Bill rightly understood.

Think about this. Neurons operate by means of neural impulses and frequency of neural firing. They can synapse laterally onto other nearby neurons, and affect the patterns of firing of those cells. This creates the possibility of both
spatial and temporal patterns occurring within clusters of cells. Some of those clusters may form what we in CSG call, somewhat sweepingly, “perceptual input functions.� (Notice, a lot of unspecified details are hidden behind our terms here.)

Now, when a reference signal comes from a higher level, exactly how is it to specify what patterns will be the standard of reference, to be reproduced by the control loops at that point, and implemented by the control loops cascading down
from that level in the hierarchy? This is no trivial problem. And the way Bill approaches it I think is quite brilliant.

Returning to p. 217 in the Memory chapter of B:CP (1st ed.), Bill introduces a key postulate:

“We will assume from now on that all reference signals are retrieved recordings of past perceptual signals.� (italics omitted)

This is the first step in handling incommensurability. If the pattern of a current perceptual signal is being compared to the pattern of a past perceptual signal, at the same point in the brain where that signal first arose, then at least
apples are being compared to apples.

Bill goes on to draw the implication of that postulate, continuing on p. 217:

“This requires giving the outputs from higher-order systems the function of address signals, whereas formerly they were reference signals. The address signals select from lower-order memory those past values of perceptual signals that
are to be recreated in present time.�

This is a part of PCT we rarely discuss. Reference signals are address signals. It is like saying, “Give me more of “_that_�.� The reference doesn’t even need to know what “that� signifies. The “address� here is a way of ensuring the
reference signal is getting to the right locale, and by taking the form of a retrieved perceptual signal, both the reference and the perception are guaranteed to be speaking a comparable “language�.

There is a final step that Bill takes, which I think is highly significant. On pp. 212-213, Bill speaks of associative addressing:

“In associative addressing, the information sent to the computing device’s address input is not a location number, but a fragment of what is recorded in one or more locations in memory� (p. 212). He continues, “Any part of the stored information
can be used as an address, a match resulting in replay of the rest of the information in that memory unit� (p. 213).

This results in a pointer (an “address�) that is already contextually relevant. In other words, “Give me the rest of “that�.�

So then, a reference signal has the following features:

a) It only has to be an address signal, essentially signaling more of “that.�

b) It is a recorded copy of what it is looking for, leading to a common “currency� for comparison between perception and reference.

c) It only needs to carry a fragment of what it is looking for, because it will result in the replay of “the rest� of what it is looking for.

This is quite an amazing solution set for keeping signals commensurate between different levels of control and perception in the brain. By giving reference signals these features back in 1973, Bill was implicitly making scientific predictions
about the brain’s wiring and its functional characteristics.

Here is where it really gets interesting.

Jeff Hawkins and his colleagues at the Redwood Neuroscience Institute have articulated arrangements of cells in the neocortex of the brain that may be able to produce the features laid out under Bill Powers’ proposal.

Hawkins’ main published work in this area (with Sandra Blakeslee) is On Intelligence (2004). His goal is to discern a neurophysiologically plausible way to understand intelligence, and how actual brains construct and carry out that process.

My reservation, as I have studied Hawkins’ proposals, is that he seems to put a lot of stock in the construct of “prediction.� The core of his theory is what he calls “the memory-prediction model� (p. 5). Nonetheless, some of what his
model uses prediction for is not that far off from what Power’s model of PCT ascribes to “control.� So there may be a family resemblance going on.

For instance, Hawkins gives an analogy for what a higher region is saying to a lower region when it sends a prediction down the hierarchy: “This is basically an instruction to you about what to look for in your own input stream� (p. 136).
That sounds almost like providing a reference for a specified perception. Or consider how Hawkins talks about motor commands, (which we in CSG consider commands for certain perceptual consequences of motor behavior): “As strange as it sounds, when your own
behavior is involved, your predictions not only precede sensations, they determine sensation. …Thinking, predicting, and doing are all part of the same unfolding of sequences moving down the cortical hierarchy� (p. 158). To speak of determining sensation
is to start to speak of control.

Be that as it may, Hawkins is one of the few to try to show the broad functional utility of the “wetware� in the brain. In other words, he tries to spell out not just what brains do, but how they may be doing it.

The part that I think relates to Bill Powers’ idea of reference signals as memory address signals is when Hawkins, in chapter 6 of On Intelligence, refers to cortical inputs that relay the “name� of a given sequence. It’s a complicated
discussion, but let me see if I can walk through it.

The physiological heart of Hawkins’ work is the neocortex, in particular the almost modular arrangement into cortical columns. His research institute has tried to distill the essence of a vast literature on the horizontal and vertical
structure and synapses among the six cellular layers of those columns. He considers the neocortical columns as the basic functional units, which construct invariances and sequential patterns out of the neural firings from a flow of sensory experience.

Let me give a flavor for what Hawkins is trying to do. In the process, I’ll raise various features that may be consistent with a PCT view of what is going on. These are just a series of quotes from Hawkins (2004), to hopefully provide
context for what I’ll present below:

** “Higher regions of your cortex are keeping track of the big picture while lower areas are actively dealing with the fast changing, small details� (p. 127).

** “All objects in your world are composed of subobjects that occur consistently together; that is the very definition of an object. When we assign a name to something, we do so because a set of features consistently travels together�
(p. 126).

** “(E)ach cortical region has a name for each sequence it knows. This “name� is a group of cells whose collective firing represents the set of objects in the sequence� (p. 129).

** “So whenever I see any of these events, I will refer to them by a common name. It is this group name, not the individual patterns, that I will pass on to higher regions of the cortex� (p. 129).

** “By collapsing predictable sequences into “named objects� at each region in our hierarchy, we achieve more and more stability the higher we go. This creates invariant representations� (p. 130).

** “The opposite effect happens as a pattern moves back down the hierarchy: stable patterns get “unfolded� into sequences� (p. 130).

** “(T)he unfolding pattern is not a rigid sequence, but the end result is the same: slow-changing, high-level patterns unfolding into faster-changing, low-level patterns� (p. 132).

In the last quotation listed above, it is worth noticing the right timing relationships, which would be needed for stable hierarchical control. Higher level patterns would change more slowly, lower level patterns would change more quickly.
Without those relative time differentiations, you cannot obtain stabilized control within a hierarchical system.

With those kinds of components, Hawkins identifies what he considers the functional job of cells at various layers of the cortical column. For instance, he notes: “Converging inputs from lower regions always arrive at layer 4 – the main
input layer� (p. 1411). And because of synapses to other layers within the column, it seems “the entire column becomes active when driven from below� (p. 148). In PCT, we might think of the columnar pattern of firing as constructing a perception out of the
ascending inputs coming from lower levels of perception in the cortex.

In a similar manner, Hawkins notes: “Layer 6 cells are the downward-projecting output cells from a cortical column and project to layer 1 in the regions hierarchically below� (p. 141). From a PCT perspective, we’re starting to see the
construction of input and output functions, including their possible locale within the cortical columns. He even makes a fascinating aside, a bit later on: “(I)n addition to projecting to lower cortical regions, layer 6 cells can send their output back into
layer 4 cells of their own column. When they do, our predictions become the input. This is what we do when daydreaming or thinking� (p. 156). Seems like we’re getting close to a physiological locale for what in PCT we have called “the imagination connection.�

Let’s shift over to aspects that may represent reference signals, or more specifically Powers’ notion of “address signals.� Hawkins speaks of “two inputs to layer 1. Higher regions of cortex spread activity across layer 1 in lower regions.
Active columns within a region also spread activity across layer 1 in the same region via the thalamus� (p. 146). He goes on to suggest the functional significance of these two inputs. “We can think of these inputs to layer 1 as the name of a song (input
from above) and where we are in a song (delayed activity from active columns in the same region)� (p. 146). This is remarkably similar to what I raised above about a reference signal being a contextually relevant pointer.

So then, there are three main streams of activation in Hawkins’ schematic layout: “converging patterns going up the cortical hierarchy, diverging patterns going down the cortical hierarchy, and a delayed feedback through the thalamus� (p.
147). He further specifies how these streams interact.

He states, “(H)alf of the input to layer 1 comes from layer 5 cells in neighboring columns and regions of the cortex. This information represents what was happening moments before� (p. 149). He also notes, “The other half of the input
to layer 1 comes from layer 6 cells in hierarchically higher regions. …It represents the name of the sequence you aree currently experiencingâ€? (p. 149). This is the portion I paraphrased above with the analogy, “Give me more of that.â€?

Hawkins’ conclusion is as follows: “Thus the information in layer 1 represents both the name of a sequence and the last item in the sequence. In this way, a particular column can be shared among many different sequences without getting
confused. Columns learn to fire in the right context and in the correct order� (p. 149). Thus, there is commensurate communication, which is not getting confused as to where and how it is talking. It uses the right name, in the right context, using a common
currency to avoid confusion.

There are a few fine points that fit into the picture. The cortical columns seem to have ways of storing the sequences they have constructed. Once a column has become active, it seems through synaptic strengthening that cells in layers
2, 3, and 5 can learn to keep firing, even when the initiating input via layer 4 is no longer active. That would suggest that they can be summoned (or retrieved), given the right naming input from above.

It appears as though the “name� is a key way that columns are communicating vertically up and down the hierarchy. Going up the hierarchy, “(w)hen a layer 4 cell fires, it is “voting� that the input fits its label� (p. 147). Its synapses
get the whole column active, and further projections continue up the hierarchy. Furthermore, there is lateral inhibition of nearby columns, to further shape and refine the name, so that higher regions do not get a jumble of possible names from below.

Going down the hierarchy, a layer 6 cell does the communicating. Hawkins suggests what layer 6 may be saying: “I speak for my region of cortex, …my job is to tell the lower regions of cortex what we think is happening. I represent our
interpretation of the worldâ€? (p. 154f.). That descending output is projected (via layer 1) to cells of layer 2 at the next lower cortical level, which “learn to be driven purely from the hierarchically higher regions of cortex. …The layer 2 cells would therefore
represent the constant name pattern from the higher region� (p. 153). Here we come incredibly close to seeing a reference signal in action, as constrained by the features that Powers laid out.

Hawkins summarizes the overall scheme as follows: “Every moment in your waking life, each region of your neocortex is comparing a set of expected columns driven from above with the set of observed columns driven from below� (p. 156). To
bring this closer in line with Perceptual Control Theory, we would call the signals from above as the “desiredâ€? input, not merely the “expectedâ€? set. To use a PCT formulation, perceptions get compared to §references. Either way the “bottom-up/top-down
matching mechanism� (p. 156) enables sending the right contextually relevant name to the next lower region of the hierarchy. That fragment of what is wanted will then be matched and unfolded into relevant patterns of what is desired, all the way down.

The striking part for me is that Jeff Hawkins’ neurophysiological way of understanding how the cortex works is quite compatible with the broad functional contours that Bill Powers laid out, three decades earlier. I have not seen indications
that Hawkins is familiar with Powers’ approach, although his institute has obviously surveyed a wide swatch of theory and research into the brain. If there is no direct familiarity, I think that would make this potential overlap between these two theories
all the more remarkable.

Comments, criticisms, and questions are welcome.

All the best,

Erling

NOTICE: This e-mail communication (including any attachments) is CONFIDENTIAL and the materials contained herein are PRIVILEGED and intended only for disclosure to or use by the person(s) listed
above. If you are neither the intended recipient(s), nor a person responsible for the delivery of this communication to the intended recipient(s), you are hereby notified that any retention, dissemination, distribution or copying of this communication is strictly
prohibited. If you have received this communication in error, please notify me immediately by using the “reply” feature or by calling me at the number listed above, and then immediately delete this message and all attachments from your computer. Thank you.

[From Erling Jorgensen (2014.08.22 12:20 EDT)]

Chad Green (2014.08.21 15:03 ET)

Erling, I wanted to let you know that I just shared your observations below
with Francis Heylighen of the Global Brain Institute
(Francis Heylighen: home page) since he had expressed a strong
interest in Hawkins’s book On Intelligence.

... [specifically -- Erling Jorgensen (2013.11.16 1100EST) CSGnet post] ...

Hello Chad,

Thank you for doing that. I'm honored. In past years I have read a lot of
Francis Heylighen's work on Principia Cybernetica Web, and been very
impressed. He is a very broad and synthetic kind of thinker. (I also
noticed, from your link above, that he is the same Myers-Briggs Personality
Type as me, INTP, so maybe that's the source of the appeal...)

The material I am preparing for the book that Warren Mansell is editing will
try to take these ideas on memory and address signals further. I try to
relate them to Dileep George's dissertation, who is a colleague of Jeff
Hawkins. George tried to lay out specific algorithms for what cells in the
cortical layers may be doing, albeit from a Bayesian predictive approach to
modeling. But I think there are a lot of parallels to be mined, and that
Hawkins' and George's approach might benefit from what PCT has to offer.
We'll see, I guess.

Thank you for your interest.

All the best,
Erling

B:CP Chapter 15: Memory & Address Signals

[From Erling Jorgensen (2013.11.16 1100EST)]

I’m a little behind in keeping up with the chapters.

There is a piece of Bill Powers’ chapter on memory that is only slowly coming into resolution for me. It strikes me as a critical piece, but I haven’t seen much discussion of it. By way of forewarning, this turned into an extended essay, touching first on Power’s understanding of reference signals, and then examining Jeff Hawkins’ treatment of the neurophysiology of the cerebral cortex of the brain.

I’m using the first edition of Behavior: The Control of Perception (1973).

On p. 217, Bill says, almost in passing:

“This, incidentally, solves the problem of translating an error in a higher-order variable into a specific value of a lower-order perception, a problem that has quietly been lurking in the background.�

I’d like to unpack what I think is going on in this passage and in the surrounding pages of B:CP.

First, what problem is being solved? And why is it a problem in the first place? To give it a name, it is the issue of commensurability. Wikipedia defines commensurability as follows: “Two concepts or things are commensurable if they are measurable or comparable by a common standard.�

It’s the problem of interface. How will two things communicate in a way that is understandable to both sides? What language or units or frames do they hold in common?

At this juncture in the history of Perceptual Control Theory, this may seem a trivial problem. After all, we have already had numerous rigorous demonstrations of principle, in the various computer simulations that show just how well a negative-feedback control architecture can operate. There are even simulations of hierarchical arrangements of control loops, where different classes of perception are being defined by the perceptual input functions, but stable control is achieved nonetheless as higher levels set reference standards for lower level loops.

In all these simulations, commensurability between the different levels is presupposed. It’s handled by the computer. The equations assign numbers to the different variables, and the computations proceed straightaway. It is why we operate with an assumption that variables perform as scalar quantities, even if they may represent a more complex form of perception.

That scalar assumption is itself based on a decision that Bill sets out on page 22 of B:CP:

“As the basic measure of nervous-system activity, therefore, I choose to use neural current, defined as the number of impulses passing through a cross section of all parallel redundant fibers in a given bundle per unit time.� (italics omitted)

Bill was an engineer, and he knew you had to pay attention to the units and definitions of the variables, so that they could be properly compared and manipulated. So again, the issue of commensurability seems to be solved, by how Bill sets out his definitions. When we lay out equations for how control loops operate, so says he, we’re talking about “neural currents.� That is the common measure that will make the different layers commensurate.

This is a credible (and insightful) way to map the concepts of engineering control theory onto the biological mechanisms of organisms with their nervous systems. Whenever we build maps, we use what can be called “convenient fictions,� which simplify but hopefully retain essential features of the territory we are trying to map.

The problem that Bill knew was “quietly lurking in the background� was that it can sometimes be hard to squeeze the underlying details into the convenient fictions we choose for our map. In the case of nervous system operation, there is a sizeable risk of incommensurability, as Bill rightly understood.

Think about this. Neurons operate by means of neural impulses and frequency of neural firing. They can synapse laterally onto other nearby neurons, and affect the patterns of firing of those cells. This creates the possibility of both spatial and temporal patterns occurring within clusters of cells. Some of those clusters may form what we in CSG call, somewhat sweepingly, “perceptual input functions.� (Notice, a lot of unspecified details are hidden behind our terms here.)

Now, when a reference signal comes from a higher level, exactly how is it to specify what patterns will be the standard of reference, to be reproduced by the control loops at that point, and implemented by the control loops cascading down from that level in the hierarchy? This is no trivial problem. And the way Bill approaches it I think is quite brilliant.

Returning to p. 217 in the Memory chapter of B:CP (1st ed.), Bill introduces a key postulate:

“We will assume from now on that all reference signals are retrieved recordings of past perceptual signals.� (italics omitted)

This is the first step in handling incommensurability. If the pattern of a current perceptual signal is being compared to the pattern of a past perceptual signal, at the same point in the brain where that signal first arose, then at least apples are being compared to apples.

Bill goes on to draw the implication of that postulate, continuing on p. 217:

“This requires giving the outputs from higher-order systems the function of address signals, whereas formerly they were reference signals. The address signals select from lower-order memory those past values of perceptual signals that are to be recreated in present time.�

This is a part of PCT we rarely discuss. Reference signals are address signals. It is like saying, “Give me more of “_that_�.� The reference doesn’t even need to know what “that� signifies. The “address� here is a way of ensuring the reference signal is getting to the right locale, and by taking the form of a retrieved perceptual signal, both the reference and the perception are guaranteed to be speaking a comparable “language�.

There is a final step that Bill takes, which I think is highly significant. On pp. 212-213, Bill speaks of associative addressing:

“In associative addressing, the information sent to the computing device’s address input is not a location number, but a fragment of what is recorded in one or more locations in memory� (p. 212). He continues, “Any part of the stored information can be used as an address, a match resulting in replay of the rest of the information in that memory unit� (p. 213).

This results in a pointer (an “address�) that is already contextually relevant. In other words, “Give me the rest of “that�.�

So then, a reference signal has the following features:

a) It only has to be an address signal, essentially signaling more of “that.�

b) It is a recorded copy of what it is looking for, leading to a common “currency� for comparison between perception and reference.

c) It only needs to carry a fragment of what it is looking for, because it will result in the replay of “the rest� of what it is looking for.

This is quite an amazing solution set for keeping signals commensurate between different levels of control and perception in the brain. By giving reference signals these features back in 1973, Bill was implicitly making scientific predictions about the brain’s wiring and its functional characteristics.

Here is where it really gets interesting.

Jeff Hawkins and his colleagues at the Redwood Neuroscience Institute have articulated arrangements of cells in the neocortex of the brain that may be able to produce the features laid out under Bill Powers’ proposal.

Hawkins’ main published work in this area (with Sandra Blakeslee) is On Intelligence (2004). His goal is to discern a neurophysiologically plausible way to understand intelligence, and how actual brains construct and carry out that process.

My reservation, as I have studied Hawkins’ proposals, is that he seems to put a lot of stock in the construct of “prediction.� The core of his theory is what he calls “the memory-prediction model� (p. 5). Nonetheless, some of what his model uses prediction for is not that far off from what Power’s model of PCT ascribes to “control.� So there may be a family resemblance going on.

For instance, Hawkins gives an analogy for what a higher region is saying to a lower region when it sends a prediction down the hierarchy: “This is basically an instruction to you about what to look for in your own input streamâ€? (p. 136). That sounds almost like providing a reference for a specified perception. Or consider how Hawkins talks about motor commands, (which we in CSG consider commands for certain perceptual consequences of motor behavior): “As strange as it sounds, when your own behavior is involved, your predictions not only precede sensations, they determine sensation. …Thinking, predicting, and doing are all part of the same unfolding of sequences moving down the cortical hierarchyâ€? (p. 158). To speak of determining sensation is to start to speak of control.

Be that as it may, Hawkins is one of the few to try to show the broad functional utility of the “wetware� in the brain. In other words, he tries to spell out not just what brains do, but how they may be doing it.

The part that I think relates to Bill Powers’ idea of reference signals as memory address signals is when Hawkins, in chapter 6 of On Intelligence, refers to cortical inputs that relay the “name� of a given sequence. It’s a complicated discussion, but let me see if I can walk through it.

The physiological heart of Hawkins’ work is the neocortex, in particular the almost modular arrangement into cortical columns. His research institute has tried to distill the essence of a vast literature on the horizontal and vertical structure and synapses among the six cellular layers of those columns. He considers the neocortical columns as the basic functional units, which construct invariances and sequential patterns out of the neural firings from a flow of sensory experience.

Let me give a flavor for what Hawkins is trying to do. In the process, I’ll raise various features that may be consistent with a PCT view of what is going on. These are just a series of quotes from Hawkins (2004), to hopefully provide context for what I’ll present below:

** “Higher regions of your cortex are keeping track of the big picture while lower areas are actively dealing with the fast changing, small details� (p. 127).

** “All objects in your world are composed of subobjects that occur consistently together; that is the very definition of an object. When we assign a name to something, we do so because a set of features consistently travels together� (p. 126).

** “(E)ach cortical region has a name for each sequence it knows. This “name� is a group of cells whose collective firing represents the set of objects in the sequence� (p. 129).

** “So whenever I see any of these events, I will refer to them by a common name. It is this group name, not the individual patterns, that I will pass on to higher regions of the cortex� (p. 129).

** “By collapsing predictable sequences into “named objects� at each region in our hierarchy, we achieve more and more stability the higher we go. This creates invariant representations� (p. 130).

** “The opposite effect happens as a pattern moves back down the hierarchy: stable patterns get “unfolded� into sequences� (p. 130).

** “(T)he unfolding pattern is not a rigid sequence, but the end result is the same: slow-changing, high-level patterns unfolding into faster-changing, low-level patterns� (p. 132).

In the last quotation listed above, it is worth noticing the right timing relationships, which would be needed for stable hierarchical control. Higher level patterns would change more slowly, lower level patterns would change more quickly. Without those relative time differentiations, you cannot obtain stabilized control within a hierarchical system.

With those kinds of components, Hawkins identifies what he considers the functional job of cells at various layers of the cortical column. For instance, he notes: “Converging inputs from lower regions always arrive at layer 4 – the main input layerâ€? (p. 141). And because of synapses to other layers within the column, it seems “the entire column becomes active when driven from belowâ€? (p. 148). In PCT, we might think of the columnar pattern of firing as constructing a perception out of the ascending inputs coming from lower levels of perception in the cortex.

In a similar manner, Hawkins notes: “Layer 6 cells are the downward-projecting output cells from a cortical column and project to layer 1 in the regions hierarchically below� (p. 141). From a PCT perspective, we’re starting to see the construction of input and output functions, including their possible locale within the cortical columns. He even makes a fascinating aside, a bit later on: “(I)n addition to projecting to lower cortical regions, layer 6 cells can send their output back into layer 4 cells of their own column. When they do, our predictions become the input. This is what we do when daydreaming or thinking� (p. 156). Seems like we’re getting close to a physiological locale for what in PCT we have called “the imagination connection.�

Let’s shift over to aspects that may represent reference signals, or more specifically Powers’ notion of “address signals.� Hawkins speaks of “two inputs to layer 1. Higher regions of cortex spread activity across layer 1 in lower regions. Active columns within a region also spread activity across layer 1 in the same region via the thalamus� (p. 146). He goes on to suggest the functional significance of these two inputs. “We can think of these inputs to layer 1 as the name of a song (input from above) and where we are in a song (delayed activity from active columns in the same region)� (p. 146). This is remarkably similar to what I raised above about a reference signal being a contextually relevant pointer.

So then, there are three main streams of activation in Hawkins’ schematic layout: “converging patterns going up the cortical hierarchy, diverging patterns going down the cortical hierarchy, and a delayed feedback through the thalamus� (p. 147). He further specifies how these streams interact.

He states, “(H)alf of the input to layer 1 comes from layer 5 cells in neighboring columns and regions of the cortex. This information represents what was happening moments beforeâ€? (p. 149). He also notes, “The other half of the input to layer 1 comes from layer 6 cells in hierarchically higher regions. …It represents the name of the sequence you are currently experiencingâ€? (p. 149). This is the portion I paraphrased above with the analogy, “Give me more of that.â€?

Hawkins’ conclusion is as follows: “Thus the information in layer 1 represents both the name of a sequence and the last item in the sequence. In this way, a particular column can be shared among many different sequences without getting confused. Columns learn to fire in the right context and in the correct order� (p. 149). Thus, there is commensurate communication, which is not getting confused as to where and how it is talking. It uses the right name, in the right context, using a common currency to avoid confusion.

There are a few fine points that fit into the picture. The cortical columns seem to have ways of storing the sequences they have constructed. Once a column has become active, it seems through synaptic strengthening that cells in layers 2, 3, and 5 can learn to keep firing, even when the initiating input via layer 4 is no longer active. That would suggest that they can be summoned (or retrieved), given the right naming input from above.

It appears as though the “name� is a key way that columns are communicating vertically up and down the hierarchy. Going up the hierarchy, “(w)hen a layer 4 cell fires, it is “voting� that the input fits its label� (p. 147). Its synapses get the whole column active, and further projections continue up the hierarchy. Furthermore, there is lateral inhibition of nearby columns, to further shape and refine the name, so that higher regions do not get a jumble of possible names from below.

Going down the hierarchy, a layer 6 cell does the communicating. Hawkins suggests what layer 6 may be saying: “I speak for my region of cortex, …my job is to tell the lower regions of cortex what we think is happening. I represent our interpretation of the worldâ€? (p. 154f.). That descending output is projected (via layer 1) to cells of layer 2 at the next lower cortical level, which “learn to be driven purely from the hierarchically higher regions of cortex. …The layer 2 cells would therefore represent the constant name pattern from the higher regionâ€? (p. 153). Here we come incredibly close to seeing a reference signal in action, as constrained by the features that Powers laid out.

Hawkins summarizes the overall scheme as follows: “Every moment in your waking life, each region of your neocortex is comparing a set of expected columns driven from above with the set of observed columns driven from belowâ€? (p. 156). To bring this closer in line with Perceptual Control Theory, we would call the signals from above as the “desiredâ€? input, not merely the “expectedâ€? set. To use a PCT formulation, perceptions get compared to §references. Either way the “bottom-up/top-down matching mechanismâ€? (p. 156) enables sending the right contextually relevant name to the next lower region of the hierarchy. That fragment of what is wanted will then be matched and unfolded into relevant patterns of what is desired, all the way down.

The striking part for me is that Jeff Hawkins’ neurophysiological way of understanding how the cortex works is quite compatible with the broad functional contours that Bill Powers laid out, three decades earlier. I have not seen indications that Hawkins is familiar with Powers’ approach, although his institute has obviously surveyed a wide swatch of theory and research into the brain. If there is no direct familiarity, I think that would make this potential overlap between these two theories all the more remarkable.

Comments, criticisms, and questions are welcome.

All the best,

Erling

  NOTICE: This e-mail communication (including any attachments) is CONFIDENTIAL and the materials contained herein are PRIVILEGED and intended only for disclosure to or use by the person(s) listed above. If you are neither the intended recipient(s), nor a person responsible for the delivery of this communication to the intended recipient(s), you are hereby notified that any retention, dissemination, distribution or copying of this communication is strictly prohibited. If you have received this communication in error, please notify me immediately by using the "reply" feature or by calling me at the number listed above, and then immediately delete this message and all attachments from your computer. Thank you.

I’m trying a different way of sending the post I just sent about Memory & Address Signals. I noticed that all the quotation marks (and there are many!) came out as a weird format, & it makes it very difficult to read. So I’ve attached the word file below, & hopefully this will be a better way to open & read it. Sorry for the confusion.

Erling

MemoryAddressCommensurate.doc (43.5 KB)

[From Rupert Young (2013.11.17 17.00 UT)]

As it happens, as I am currently reading Hawkins' book I was in the midst of formulating a couple of posts about it when yours popped up.

I too was struck by a number of similarities with PCT in the language used by Hawkins. I'll continue with my posts as they cover specific issues slightly different to the points you make, though the one on predictions is related to what you discuss here.

On your point about commensurability, it seems that what we are talking about here is representation, in that do neural entities represent something in the world we recognise and do different neural entities that communicate with each other represent the same thing. Personally I am very reticent about talking about representation when it comes to neural systems as I think it is, as you say a "convenient fiction" to help us formulate and understand ideas about how systems operate. But it is using language that is recognisable by us but not actually properties of the actual systems we are attempting to describe. Living systems evolved to control neural signals that enabled them to survive. Actual systems don't "do" representation they do control. In which case is the issue of commensurability a problem of our understanding of these systems rather than a problem with how they are able to operate? Perhaps that is what you are saying.

Generally, about Hawkins' "theory", it seems to me to be very speculative about what is going on is the cortical systems, based upon a presumption of the memory-prediction model. And, from what I have read so far (up to p96), there doesn't appear to be any formal definition of the theory, as there is with PCT and the control system equations.

Also I don't see that his argument in favour of prediction is valid as he uses the term in two different senses. On one hand he talks about it in a very technical sense of the actual operation of the cortex being to compare patterns predicted by bayesian probability with sensed patterns. On the other hand he talks about prediction in an everyday sense, such as being able to predict what his wife is going to say in a specific context. His conclusion, as he appears to be making, that the latter proves the former just doesn't hold up to scrutiny.

Regards,
Rupert

···

On 16/11/2013 19:19, Erling Jorgensen wrote:

B:CP Chapter 15: Memory & Address Signals

Â

[From Erling Jorgensen (2013.11.16 1100EST)] Â

Â

I’m a little behind in keeping up with the chapters.Â

Â

There is a piece of Bill Powers’ chapter on memory that is only slowly coming into resolution for me. It strikes me as a critical piece, but I haven’t seen much discussion of it. By way of forewarning, this turned into an extended essay, touching first on Power’s understanding of reference signals, and then examining Jeff Hawkins’ treatment of the neurophysiology of the cerebral cortex of the brain.Â

Â

I’m using the first edition of Behavior: The Control of Perception (1973). Â

Â

On p. 217, Bill says, almost in passing: Â

“This, incidentally, solves the problem of translating an error in a higher-order variable into a specific value of a lower-order perception, a problem that has quietly been lurking in the background.â€?Â

Â

I’d like to unpack what I think is going on in this passage and in the surrounding pages of B:CP.Â

Â

First, what problem is being solved? And why is it a problem in the first place? To give it a name, it is the issue of commensurability. Wikipedia defines commensurability as follows: “Two concepts or things are commensurable if they are measurable or comparable by a common standard.â€?Â

Â

It’s the problem of interface. How will two things communicate in a way that is understandable to both sides? What language or units or frames do they hold in common?Â

Â

At this juncture in the history of Perceptual Control Theory, this may seem a trivial problem. After all, we have already had numerous rigorous demonstrations of principle, in the various computer simulations that show just how well a negative-feedback control architecture can operate. There are even simulations of hierarchical arrangements of control loops, where different classes of perception are being defined by the perceptual input functions, but stable control is achieved nonetheless as higher levels set reference standards for lower level loops.Â

Â

In all these simulations, commensurability between the different levels is presupposed. It’s handled by the computer. The equations assign numbers to the different variables, and the computations proceed straightaway. It is why we operate with an assumption that variables perform as scalar quantities, even if they may represent a more complex form of perception.Â

Â

That scalar assumption is itself based on a decision that Bill sets out on page 22 of B:CP:Â

“As the basic measure of nervous-system activity, therefore, I choose to use neural current, defined as the number of impulses passing through a cross section of all parallel redundant fibers in a given bundle per unit time.â€? (italics omitted)Â

Â

Bill was an engineer, and he knew you had to pay attention to the units and definitions of the variables, so that they could be properly compared and manipulated. So again, the issue of commensurability seems to be solved, by how Bill sets out his definitions. When we lay out equations for how control loops operate, so says he, we’re talking about “neural currents.â€? That is the common measure that will make the different layers commensurate.Â

Â

This is a credible (and insightful) way to map the concepts of engineering control theory onto the biological mechanisms of organisms with their nervous systems. Whenever we build maps, we use what can be called “convenient fictions,â€? which simplify but hopefully retain essential features of the territory we are trying to map.Â

Â

The problem that Bill knew was “quietly lurking in the backgroundâ€? was that it can sometimes be hard to squeeze the underlying details into the convenient fictions we choose for our map. In the case of nervous system operation, there is a sizeable risk of incommensurability, as Bill rightly understood.Â

Â

Think about this. Neurons operate by means of neural impulses and frequency of neural firing. They can synapse laterally onto other nearby neurons, and affect the patterns of firing of those cells. This creates the possibility of both spatial and temporal patterns occurring within clusters of cells. Some of those clusters may form what we in CSG call, somewhat sweepingly, “perceptual input functions.â€? (Notice, a lot of unspecified details are hidden behind our terms here.)Â

Â

Now, when a reference signal comes from a higher level, exactly how is it to specify what patterns will be the standard of reference, to be reproduced by the control loops at that point, and implemented by the control loops cascading down from that level in the hierarchy? This is no trivial problem.  And the way Bill approaches it I think is quite brilliant.Â

Â

Returning to p. 217 in the Memory chapter of B:CP (1st ed.), Bill introduces a key postulate: Â

“We will assume from now on that all reference signals are retrieved recordings of past perceptual signals.� (italics omitted)

Â

This is the first step in handling incommensurability. If the pattern of a current perceptual signal is being compared to the pattern of a past perceptual signal, at the same point in the brain where that signal first arose, then at least apples are being compared to apples.Â

Â

Bill goes on to draw the implication of that postulate, continuing on p. 217: Â

“This requires giving the outputs from higher-order systems the function of address signals, whereas formerly they were reference signals. The address signals select from lower-order memory those past values of perceptual signals that are to be recreated in present time.â€?Â

Â

This is a part of PCT we rarely discuss. Reference signals are address signals. It is like saying, “Give me more of “_that_â€?.â€? The reference doesn’t even need to know what “thatâ€? signifies. The “addressâ€? here is a way of ensuring the reference signal is getting to the right locale, and by taking the form of a retrieved perceptual signal, both the reference and the perception are guaranteed to be speaking a comparable “languageâ€?.Â

Â

There is a final step that Bill takes, which I think is highly significant. On pp. 212-213, Bill speaks of associative addressing:Â

“In associative addressing, the information sent to the computing device’s address input is not a location number, but a fragment of what is recorded in one or more locations in memoryâ€? (p. 212). He continues, “Any part of the stored information can be used as an address, a match resulting in replay of the rest of the information in that memory unitâ€? (p. 213).Â

Â

This results in a pointer (an “addressâ€?) that is already contextually relevant. In other words, “Give me the _rest_ of “thatâ€?.â€?Â

Â

So then, a reference signal has the following features:Â

a) It only has to be an address signal, essentially signaling more of “that.â€?Â

b) It is a recorded copy of what it is looking for, leading to a common “currencyâ€? for comparison between perception and reference.Â

c) It only needs to carry a fragment of what it is looking for, because it will result in the replay of “the restâ€? of what it is looking for.Â

Â

This is quite an amazing solution set for keeping signals commensurate between different levels of control and perception in the brain. By giving reference signals these features back in 1973, Bill was implicitly making scientific predictions about the brain’s wiring and its functional characteristics.Â

Â

Here is where it really gets interesting.Â

Â

Jeff Hawkins and his colleagues at the Redwood Neuroscience Institute have articulated arrangements of cells in the neocortex of the brain that may be able to produce the features laid out under Bill Powers’ proposal.Â

Â

Hawkins’ main published work in this area (with Sandra Blakeslee) is _On Intelligence_ (2004). His goal is to discern a neurophysiologically plausible way to understand intelligence, and how actual brains construct and carry out that process.Â

Â

My reservation, as I have studied Hawkins’ proposals, is that he seems to put a lot of stock in the construct of “prediction.â€? The core of his theory is what he calls “the memory-prediction modelâ€? (p. 5). Nonetheless, some of what his model uses prediction for is not that far off from what Power’s model of PCT ascribes to “control.â€? So there may be a family resemblance going on.Â

Â

For instance, Hawkins gives an analogy for what a higher region is saying to a lower region when it sends a prediction down the hierarchy: “This is basically an instruction to you about what to look for in your own input streamâ€? (p. 136). That sounds almost like providing a reference for a specified perception. Or consider how Hawkins talks about motor commands, (which we in CSG consider commands for certain perceptual consequences of motor behavior): “As strange as it sounds, when your own behavior is involved, your predictions not only precede sensations, they determine sensation. …Thinking, predicting, and doing are all part of the same unfolding of sequences moving down the cortical hierarchyâ€? (p. 158). To speak of determining sensation is to start to speak of control.Â

Â

Be that as it may, Hawkins is one of the few to try to show the broad functional utility of the “wetwareâ€? in the brain. In other words, he tries to spell out not just what brains do, but how they may be doing it.Â

Â

The part that I think relates to Bill Powers’ idea of reference signals as memory address signals is when Hawkins, in chapter 6 of _On Intelligence_, refers to cortical inputs that relay the “nameâ€? of a given sequence. It’s a complicated discussion, but let me see if I can walk through it.Â

Â

The physiological heart of Hawkins’ work is the neocortex, in particular the almost modular arrangement into cortical columns. His research institute has tried to distill the essence of a vast literature on the horizontal and vertical structure and synapses among the six cellular layers of those columns. He considers the neocortical columns as the basic functional units, which construct invariances and sequential patterns out of the neural firings from a flow of sensory experience.Â

Â

Let me give a flavor for what Hawkins is trying to do. In the process, I’ll raise various features that may be consistent with a PCT view of what is going on. These are just a series of quotes from Hawkins (2004), to hopefully provide context for what I’ll present below:Â

Â

** “Higher regions of your cortex are keeping track of the big picture while lower areas are actively dealing with the fast changing, small detailsâ€? (p. 127).Â

** “All objects in your world are composed of subobjects that occur consistently together; that is the very definition of an object. When we assign a name to something, we do so because a set of features consistently travels togetherâ€? (p. 126).Â

** “(E)ach cortical region has a name for each sequence it knows. This “nameâ€? is a group of cells whose collective firing represents the set of objects in the sequenceâ€? (p. 129).Â

** “So whenever I see any of these events, I will refer to them by a common name. It is this group name, not the individual patterns, that I will pass on to higher regions of the cortexâ€? (p. 129).Â

** “By collapsing predictable sequences into “named objectsâ€? at each region in our hierarchy, we achieve more and more stability the higher we go. This creates invariant representationsâ€? (p. 130).Â

** “The opposite effect happens as a pattern moves back down the hierarchy: stable patterns get “unfoldedâ€? into sequencesâ€? (p. 130).Â

** “(T)he unfolding pattern is not a rigid sequence, but the end result is the same: slow-changing, high-level patterns unfolding into faster-changing, low-level patternsâ€? (p. 132).Â

Â

In the last quotation listed above, it is worth noticing the right timing relationships, which would be needed for stable hierarchical control. Higher level patterns would change more slowly, lower level patterns would change more quickly. Without those relative time differentiations, you cannot obtain stabilized control within a hierarchical system.Â

Â

With those kinds of components, Hawkins identifies what he considers the functional job of cells at various layers of the cortical column. For instance, he notes: “Converging inputs from lower regions always arrive at layer 4 – the main input layerâ€? (p. 141). And because of synapses to other layers within the column, it seems “the entire column becomes active when driven from belowâ€? (p. 148). In PCT, we might think of the columnar pattern of firing as constructing a perception out of the ascending inputs coming from lower levels of perception in the cortex.Â

Â

In a similar manner, Hawkins notes: “Layer 6 cells are the downward-projecting output cells from a cortical column and project to layer 1 in the regions hierarchically belowâ€? (p. 141). From a PCT perspective, we’re starting to see the construction of input and output functions, including their possible locale within the cortical columns. He even makes a fascinating aside, a bit later on: “(I)n addition to projecting to lower cortical regions, layer 6 cells can send their output back into layer 4 cells of their own column. When they do, our predictions become the input. This is what we do when daydreaming or thinkingâ€? (p. 156). Seems like we’re getting close to a physiological locale for what in PCT we have called “the imagination connection.â€?Â

Â

Let’s shift over to aspects that may represent reference signals, or more specifically Powers’ notion of “address signals.â€? Hawkins speaks of “two inputs to layer 1. Higher regions of cortex spread activity across layer 1 in lower regions. Active columns within a region also spread activity across layer 1 in the same region via the thalamusâ€? (p. 146). He goes on to suggest the functional significance of these two inputs. “We can think of these inputs to layer 1 as the name of a song (input from above) and where we are in a song (delayed activity from active columns in the same region)â€? (p. 146). This is remarkably similar to what I raised above about a reference signal being a contextually relevant pointer.Â

Â

So then, there are three main streams of activation in Hawkins’ schematic layout: “converging patterns going up the cortical hierarchy, diverging patterns going down the cortical hierarchy, and a delayed feedback through the thalamusâ€? (p. 147). He further specifies how these streams interact.Â

Â

He states, “(H)alf of the input to layer 1 comes from layer 5 cells in neighboring columns and regions of the cortex. This information represents what was happening moments beforeâ€? (p. 149). He also notes, “The other half of the input to layer 1 comes from layer 6 cells in hierarchically higher regions. …It represents the name of the sequence you are currently experiencingâ€? (p. 149). This is the portion I paraphrased above with the analogy, “Give me more of _that_.â€?Â

Â

Hawkins’ conclusion is as follows: “Thus the information in layer 1 represents both the name of a sequence and the last item in the sequence. In this way, a particular column can be shared among many different sequences without getting confused. Columns learn to fire in the right context and in the correct orderâ€? (p. 149). Thus, there is commensurate communication, which is not getting confused as to where and how it is talking. It uses the right name, in the right context, using a common currency to avoid confusion.Â

Â

There are a few fine points that fit into the picture. The cortical columns seem to have ways of storing the sequences they have constructed. Once a column has become active, it seems through synaptic strengthening that cells in layers 2, 3, and 5 can learn to keep firing, even when the initiating input via layer 4 is no longer active. That would suggest that they can be summoned (or retrieved), given the right naming input from above.Â

Â

It appears as though the “nameâ€? is a key way that columns are communicating vertically up and down the hierarchy. Going up the hierarchy, “(w)hen a layer 4 cell fires, it is “votingâ€? that the input fits its labelâ€? (p. 147). Its synapses get the whole column active, and further projections continue up the hierarchy. Furthermore, there is lateral inhibition of nearby columns, to further shape and refine the name, so that higher regions do not get a jumble of possible names from below.Â

Â

Going down the hierarchy, a layer 6 cell does the communicating. Hawkins suggests what layer 6 may be saying: “I speak for my region of cortex, …my job is to tell the lower regions of cortex what we think is happening. I represent our interpretation of the worldâ€? (p. 154f.). That descending output is projected (via layer 1) to cells of layer 2 at the next lower cortical level, which “learn to be driven purely from the hierarchically higher regions of cortex. …The layer 2 cells would therefore represent the constant name pattern from the higher regionâ€? (p. 153). Here we come incredibly close to seeing a reference signal in action, as constrained by the features that Powers laid out.Â

Â

Hawkins summarizes the overall scheme as follows: “Every moment in your waking life, each region of your neocortex is comparing a set of expected columns driven from above with the set of observed columns driven from belowâ€? (p. 156). To bring this closer in line with Perceptual Control Theory, we would call the signals from above as the “desiredâ€? input, not merely the “expectedâ€? set. To use a PCT formulation, perceptions get compared to (p)references. Either way the “bottom-up/top-down matching mechanismâ€? (p. 156) enables sending the right contextually relevant name to the next lower region of the hierarchy. That fragment of what is wanted will then be matched and unfolded into relevant patterns of what is desired, all the way down.Â

Â

The striking part for me is that Jeff Hawkins’ neurophysiological way of understanding how the cortex works is quite compatible with the broad functional contours that Bill Powers laid out, three decades earlier. I have not seen indications that Hawkins is familiar with Powers’ approach, although his institute has obviously surveyed a wide swatch of theory and research into the brain. If there is no direct familiarity, I think that would make this potential overlap between these two theories all the more remarkable.Â

Â

Comments, criticisms, and questions are welcome.Â

Â

All the best,

Erling

Â

NOTICE: This e-mail communication (including any attachments) is CONFIDENTIAL and the materials contained herein are PRIVILEGED and intended only for disclosure to or use by the person(s) listed above. If you are neither the intended recipient(s), nor a person responsible for the delivery of this communication to the intended recipient(s), you are hereby notified that any retention, dissemination, distribution or copying of this communication is strictly prohibited. If you have received this communication in error, please notify me immediately by using the "reply" feature or by calling me at the number listed above, and then immediately delete this message and all attachments from your computer. Thank you.

[Martin Taylor 2013.11.17.09.13]

            B:CP Chapter 15:

Memory & Address Signals

Â

            [From Erling

Jorgensen (2013.11.16 1100EST)] Â

Â

Â

            There is a piece of

Bill Powers’ chapter on memory that is only slowly
coming into resolution for me. It strikes me as a critical piece, but
I haven’t seen much discussion of it. By way of
forewarning, this turned into an extended essay,
touching first on Power’s understanding of reference
signals, and then examining Jeff Hawkins’ treatment of
the neurophysiology of the cerebral cortex of the brain.Â

Erling discusses a part of Bill's model that I had intended to bring

up, but I had intended to bring it up only after the end of the book
series, because David asked me to withhold comment on the book until
then. But since Erling has provided physiological support for Bill’s
ideas, I think it appropriate to consider other implications now,
rather than wait until all the chapters have been treated. And
Erling’s post already started a thread, which could be considered a
distraction from continuing through the book. So if this is
inappropriate, I apologize, and will delay further messages on this
theme.

Erling used the first edition of B:CP. I am using the second. In

particular, I refer to Figure 15.3 on page 223.

Â

Â

            Now, when a

reference signal comes from a higher level, exactly how
is it to specify what patterns will be the standard of
reference, to be reproduced by the control loops at that
point, and implemented by the control loops cascading
down from that level in the hierarchy? This is no trivial
problem. Â And
the way Bill approaches it I think is quite brilliant.Â

Â

            Returning to p. 217

in the Memory chapter of B:CP (1st ed.), Bill
introduces a key postulate: Â

            “We will assume from

now on that all reference signals are retrieved
recordings of past perceptual signals.� (italics
omitted)

Â

            This is the first

step in handling incommensurability. If the pattern of
a current perceptual signal is being compared to the
pattern of a past perceptual signal, at the same point
in the brain where that signal first arose, then at
least apples are being compared to apples.Â

Â

            Bill goes on to draw

the implication of that postulate, continuing on p. 217:
Â

            “This requires

giving the outputs from higher-order systems the
function of address signals, whereas formerly they were
reference signals.Â
The
address signals select from lower-order memory those
past values of perceptual signals that are to be
recreated in present time.â€?Â

Â

            This is a part of

PCT we rarely discuss. Â
Reference signals are address signals. It is like saying,
“Give me more of “_that_�.� The reference doesn’t even need to know
what “thatâ€? signifies. Â
The “address� here is a way of ensuring the
reference signal is getting to the right locale, and by
taking the form of a retrieved perceptual signal, both
the reference and the perception are guaranteed to be
speaking a comparable “languageâ€?.Â

Â

            There is a final

step that Bill takes, which I think is highly
significant. On
pp. 212-213, Bill speaks of associative addressing:Â

            “In associative

addressing, the information sent to the computing
device’s address input is not a location number, but a
fragment of what is recorded in one or more locations in
memoryâ€? (p. 212).Â
He
continues, “Any part of the stored information can be
used as an address, a match resulting in replay of the
rest of the information in that memory unitâ€? (p. 213).Â

Â

            This results in a

pointer (an “address�) that is already contextually
relevant. In
other words, “Give me the rest of “thatâ€?.â€?Â

Â

            So then, a reference

signal has the following features:Â

a)Â It only has to be
an address signal, essentially signaling more of “that.â€?Â

b)Â It is a recorded
copy of what it is looking for, leading to a common
“currency� for comparison between perception and
reference.Â

c)Â It only needs to
carry a fragment of what it is looking for, because it
will result in the replay of “the rest� of what it is
looking for.Â

Â

            This is quite an

amazing solution set for keeping signals commensurate
between different levels of control and perception in
the brain. By
giving reference signals these features back in 1973,
Bill was implicitly making scientific predictions about
the brain’s wiring and its functional characteristics.Â

To begin with a comment on what Erling says, I do not see a problem

of incommensurability anywhere in the PCT hierarchy, because all the
signals are either neural firings or variations of potential within
a neural cell body. What differs from one level of the hierarchy to
another is the relation an Analyst sees between the Analyst’s
perception of the person’s environment and the Analyst’s model of
the hierarchy. Internal to the person’s hierarchy there are only
neural (electrical and chemical) fluctuations. Incommensurability
doesn’t enter the picture, so far as I can see.

I am, however, interested by the physiological part of Erling's

essay, on which I will not comment because of ignorance.

Now, about Figure 15.3. This figure is a picture of Bill's

conception of a complete elementary control unit (ECU), including
the “imagination loop” and the construction of the reference value
from an addressed memory of some prior perceptual value. The output
function and the input function at the bottom of the figure show
distribution of the signals to and acceptance from many lower-level
ECUs, but at the top of the picture, the perceptual signal and the
incoming reference signal are shown as going to and coming from only
one ECU at the next higher level. The distribution to and from many
higher-level ECUs is taken for granted.

Ideally, the top of the figure should show branching arrows in the

same way as they are shown at the bottom of the figure. Here are
some duplicated sketches of what the top-right of Fig 15.3 should
look like if it were consistent with the bottom-right of the figure.
I show a row of memory “cells” that are parts of five distinct ECUs.
You should imagine that there are hundreds (or millions) of similar
structures.

Re BCP Chapter 15 Memory & A.jpg

···

http://ecacs.net/discus/messages/192/408.html?1298907804

[From Rick Marken (2013.11.19.1605)]

Re BCP Chapter 15 Memory & A.jpg

···

[Martin Taylor 2013.11.17.09.13]

Erling discusses a part of Bill's model that I had intended to bring

up, but I had intended to bring it up only after the end of the book
series, because David asked me to withhold comment on the book until
then. But since Erling has provided physiological support for Bill’s
ideas, I think it appropriate to consider other implications now,
rather than wait until all the chapters have been treated. And
Erling’s post already started a thread, which could be considered a
distraction from continuing through the book. So if this is
inappropriate, I apologize, and will delay further messages on this
theme.

RM: No need to apologize Martin; this is a very pertinent post to a chapter we’ve already discussed and though I haven’t had a chance yet to read your post in detail it looks brilliant. Looking at the multiple higher level outputs as the vector address for a lower level reference signal is very clever!

Best

Rick

Erling used the first edition of B:CP. I am using the second. In

particular, I refer to Figure 15.3 on page 223.

            Now, when a

reference signal comes from a higher level, exactly how
is it to specify what patterns will be the standard of
reference, to be reproduced by the control loops at that
point, and implemented by the control loops cascading
down from that level in the hierarchy? This is no trivial
problem. And
the way Bill approaches it I think is quite brilliant.

            Returning to p. 217

in the Memory chapter of B:CP (1st ed.), Bill
introduces a key postulate:

            “We will assume from

now on that all reference signals are retrieved
recordings of past perceptual signals.” (italics
omitted)

            This is the first

step in handling incommensurability. If the pattern of
a current perceptual signal is being compared to the
pattern of a past perceptual signal, at the same point
in the brain where that signal first arose, then at
least apples are being compared to apples.

            Bill goes on to draw

the implication of that postulate, continuing on p. 217:

            “This requires

giving the outputs from higher-order systems the
function of address signals, whereas formerly they were
reference signals.
The
address signals select from lower-order memory those
past values of perceptual signals that are to be
recreated in present time.”

            This is a part of

PCT we rarely discuss.
Reference signals are address signals. It is like saying,
“Give me more of “_that_”.” The reference doesn’t even need to know
what “that” signifies.
The “address” here is a way of ensuring the
reference signal is getting to the right locale, and by
taking the form of a retrieved perceptual signal, both
the reference and the perception are guaranteed to be
speaking a comparable “language”.

            There is a final

step that Bill takes, which I think is highly
significant. On
pp. 212-213, Bill speaks of associative addressing:

            “In associative

addressing, the information sent to the computing
device’s address input is not a location number, but a
fragment of what is recorded in one or more locations in
memory” (p. 212).
He
continues, “Any part of the stored information can be
used as an address, a match resulting in replay of the
rest of the information in that memory unit” (p. 213).

            This results in a

pointer (an “address”) that is already contextually
relevant. In
other words, “Give me the rest of “that”.”

            So then, a reference

signal has the following features:

a) It only has to be
an address signal, essentially signaling more of “that.”

b) It is a recorded
copy of what it is looking for, leading to a common
“currency” for comparison between perception and
reference.

c) It only needs to
carry a fragment of what it is looking for, because it
will result in the replay of “the rest” of what it is
looking for.

            This is quite an

amazing solution set for keeping signals commensurate
between different levels of control and perception in
the brain. By
giving reference signals these features back in 1973,
Bill was implicitly making scientific predictions about
the brain’s wiring and its functional characteristics.

To begin with a comment on what Erling says, I do not see a problem

of incommensurability anywhere in the PCT hierarchy, because all the
signals are either neural firings or variations of potential within
a neural cell body. What differs from one level of the hierarchy to
another is the relation an Analyst sees between the Analyst’s
perception of the person’s environment and the Analyst’s model of
the hierarchy. Internal to the person’s hierarchy there are only
neural (electrical and chemical) fluctuations. Incommensurability
doesn’t enter the picture, so far as I can see.

I am, however, interested by the physiological part of Erling's

essay, on which I will not comment because of ignorance.

Now, about Figure 15.3. This figure is a picture of Bill's

conception of a complete elementary control unit (ECU), including
the “imagination loop” and the construction of the reference value
from an addressed memory of some prior perceptual value. The output
function and the input function at the bottom of the figure show
distribution of the signals to and acceptance from many lower-level
ECUs, but at the top of the picture, the perceptual signal and the
incoming reference signal are shown as going to and coming from only
one ECU at the next higher level. The distribution to and from many
higher-level ECUs is taken for granted.

Ideally, the top of the figure should show branching arrows in the

same way as they are shown at the bottom of the figure. Here are
some duplicated sketches of what the top-right of Fig 15.3 should
look like if it were consistent with the bottom-right of the figure.
I show a row of memory “cells” that are parts of five distinct ECUs.
You should imagine that there are hundreds (or millions) of similar
structures.

The "Memory" cell of each control unit is addressed by a _pattern_

of values of the outputs from higher-level ECUs, not by the value of
just one higher level output. The set of (in this case five)
retrieved reference values is also a pattern (but only in the eyes
of the Analyst). The pattern of reference values is not a
prescription for routinized action. It simply specifies to the lower
level systems that they should produce actions that would in the
current disturbance situation recreate a pattern of perceptions that
once before had existed – perhaps not all simultaneously, though we
will address that question below.

A pattern (a vector) of Address Signals such as {2.3, 1.7, 4.0,

15.2, …} is a much more specific way to address an associative
memory than a single neural current value would be. The reference
value produced for that particular lower-level ECU is much easier to
distinguish from the reference value associated with other patterns
than would be the case if the Address were to come from just one
signal value, even if that signal value were to be produced by some
function of many higher-level output values. Furthermore, the
reference value produced by a patterned Address Signal vector is
context-dependent. Much of that context is also shared by the memory
cells of other ECUs, which allows them as a group to request
perceptual values that have in the past been successfully
coordinated. One of the false issues that have been claimed to
demonstrate the inadequacy of PCT has been that the different
controlled perceptions are uncoordinated whereas we observe that in
the process of maturing, most people’s actions become ever more
coordinated. Bill’s proposal accomplishes this coordination, though
in B:CP he does not follow that train of thought very far, if at
all.

The more difficult issue is how the stored memories are selected out

of the continuous flow of varying perceptions. Bill says (p224)
“Something special must be required to result in temporally
connected address signals. Nothing in the model requires this to
happen.” He then goes on to examine the consequences of assuming it
does not happen.

So far in this message, I have just been interpreting what Bill

wrote, but now I’m going to go out on a limb, and suggest that when
we are actively controlling in the environment, the temporal
connection exists, because “now” consists of whatever perceptions of
the environment (and its history) currently influence the perceptual
side of the control hierarchy, and therefore the corresponding error
signals are temporally coordinated. But when we are operating in
imagination mode, no such temporal coordination of the recovery
address signals is imposed. Bits and pieces of perceptions that
occured at differnt times and in different contexts could well be
retrieved together, as Bill goes on to discuss (p 225ff).

Now the question is: which of the continuum of perceptual

possibilities is likely to be stored in these associative memories.
I’m going further out on my limb, to suggest it might be related to
a common feature of neural firing – adaptation, fatigue, or
whatever you might like to call it. Sensors and synapses both work
by changing a chemical source supply of some kind into electrical
energy. If the input state continues unchanged, the supply of source
chemical gets depleted, and the sensor or postsynaptic part of the
synapse produces less electrical effect (fewer impulses, or less
influence on the downstream neuron to produce impulses). A retinal
sensor on which the light level suddenly increases has a brief surge
in its firing rate, and when the light level decreases, it has a
brief depression in its resting firing rate. The same is true for a
synapse; if the presynaptic neuron keeps firing at a regular rate
fast enough that the synapse’s chemical resources cannot be restored
between firings, the postsynaptic response will become depressed.

Everywhere in the system, for the most part, changes are what

matter. The perception of intensity, whether of an environmental
variable or of any higher-level perception, is an interpretation of
some kind, such as an integration by the receiving part of the
neural system. Velocity would seem to be a primary variable,
possibly at every level of the hierarchy. In fact, when I was
rereading Chapter 10, I was wondering why “events” might be treated
as a separate level, if events (changes of state) are the important
output from every level to the next higher level.

Many times in the early chapters, Bill refers to the fact that there

are many lateral interconnections in the perceiving system.
Nowadays, 40 years after B:CP, much more is known about these, but
even then we knew that they resulted in spatial differential
perceptual variables produced by on-centre-off-surround (and the
inverse), light-dark and dark-light edges, and so on. Not only is
the primary sensory input differential in time, it is differential
in space. I suspect that the same is true at most levels of the
hierarchy. When we look at a forested landscape, we can see trees,
but we usually don’t. Instead we see a forest. But we do see the
tall tree that stands out, or the maple among cedars. We see
configuration changes, spatially and temporally. We don’t (but can)
hear the fan noise constantly in the background, but we do perceive
consciously the few seconds of fan noise before the fan is
unexpectedly turned off by someone else.

What this is leading up to is the suggestion that possible the

memories that are stored are biased in favour of perceptions at the
edge – at moments or places of change (spatial or temporal “events”
at that level); perhaps the change is not in the perceptual value
being stored, but in the error signal of the ECU that controls this
perception or that. At a global level, those of us who were
sufficiently mature in 1963 may well remember not only where we were
when we heard that Kennedy had been shot, but half a century later
can picture our surroundings and remember what we were doing. But we
don’t remember what we were doing an hour or a day earlier or later.
For many of us, that news produced error signals at a rather high
level in the hierarchy, which would have propagated real or imagined
output signals down, causing memory storage all through the
hierarchy.

At a more microscopic level, both losing or, through reorganization,

regaining control of a perception in a specific context would be
accompanied by a significant change in either the value or the rate
of change of the error variable in a control system, and I propose
that such a change would be an appropriate moment to store a
perception in its pattern context, as either a situation to be
recovered or a situation to be avoided. In other words, coordinated
perceptions would be likely to be stored in the same contextual
moment. The storage address is the retrieval address, and because
the perceptual environment exists in the perceptual “now”, what is
stored is likely to be fairly widely coordinated over time.
Similarly, if there are differences in the errors for neighbouring
perceptions (such as, say at a very low level the colour of a patch,
or at a very high level the cultural complex of practices of two
groups), those differences might induce perceptual storage, but
independently of temporal variation.

There's lots more that could be said and studied about the

implications of deriving reference values from the addressed
associative memory of past perceptual values, but the summary points
are:

    1. Bill's proposal resolves contextual dependency and

coordination issues that exist when reference values are taken to be
a single-valued function of a defined set of higher-level output
values.

    2. The stored pattern of memories that could be evoked is

likely, but not certainly, coordinated across time, at least for
localized regions of perceptual space.

    3. Most likely to be stored for later retrieval are the

perceptual values at changes across space or time.

 --------



These are ideas that I have been mulling over for several years (see

http://ecacs.net/discus/messages/192/408.html?1298907804 for
a short three-year-old discussion of the topic), but they have been
jogged and crystallized by re-reading B:CP after perhaps 20 years
since I last read it.

I'm sorry to have sent this post before we finish going through the

book. I would have waited as David had requested, but the ideas
seemed to tie in so well with Erling’s message and the ensuing
thread that I resolved my conflict in favour of posting now. The
message is long, but I think it is also a much too cursory
discussion of what is actually quite a deep subject. I hope it is at
least partially intelligible and makes some sense if you do
understand it.

Martin


Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[From Erling Jorgensen (2013.11.24 1955EST)]

Martin Taylor 2013.11.17.09.13

Hi Martin,

I was intrigued by your post about a Reference vector, which provides a
context-relevant Address Signal to lower level memory. Some of this will
be paraphrasing your ideas to see if I am understanding what you are
getting at. I am also following with interest your more recent post
integrating emotion into this memory addressing system, but I haven't fully
digested that line of argument.

I'll begin, however, with a rebuttal-type response on this issue of
commensurability, in your reply to my post of [Erling Jorgensen (2013.11.16
1100EST)].

To begin with a comment on what Erling says, I do not see a problem of
incommensurability anywhere in the PCT hierarchy, because all the signals
are either neural firings or variations of potential within a neural cell
body.

Yes, both you and Rick Marken (2013.11.19.0930) hold the view that Powers'
specification of "neural current" creates a common unit among neural
control systems and thus solves the problem of incommensurability.

However, Bill himself seems to have thought there was still "a problem that
has quietly been lurking in the background" (B:CP, 1st ed., p. 217). While
the common currency of neural currents was necessary, it seems not to have
been sufficient, to make sure reference signals are talking with comparable
perceptual signals within Bill's model.

In the quotations from B:CP chapter 15 that I laid out in my previous post
on Memory & Address Signals, I list what Bill seemed to think were
additional constraints on how reference signals were to operate in his
model.

Powers says quite plainly, "Memory, in order to be useful, must be under
control of some orderly addressing mechanism. That is the principle I wish
to communicate" (p. 216, emphasis omitted). So here he is saying that, in
addition to having a comparable current, reference signals must specify a
comparable locale. The address must be right.

If I am looking to produce the right visual sequences of letters and words
on this screen as I type, I don't think it will work if the reference
specifications are being sent to my auditory cortex. Yes, sound is part of
the overall context of typing, and so if I am not hearing the expected
tapping of the keyboard keys, I will think something is wrong. But those
taps in themselves do not differentiate the letters sufficiently for me to
get the right pattern of words on the screen. The reference signals must
be going to the proper address, if they are to function properly.

In addition, Bill made the point that reference and perceptual signals must
also share a comparable pattern. This is spelled out when he says, "all
reference signals are retrieved recordings of past perceptual signals" (p.
217). Like-to-like seems to be the principle here.

So then, in order to avoid a problem of incommensurability between
reference signals and perceptual signals, the following conditions must be
met:

1) Comparable Current

2) Comparable Locale

3) Comparable Pattern (or fragment of a pattern)

It seems, Martin, that you are adding a fourth possible condition:
Comparable Context. And here I'm shifting from rebuttal to review, to
check my understanding of your ideas.

You make the point that a memory address constructed out of a vector of
higher-level outputs is a more stable reference than a single-valued
output. The address becomes much more specific as a result. To note a
trivial analogy, this is why postal addresses don't just have a house (or
apartment) number. That would lead to too many false negatives. The post
office also wants a street name, and a city name, and a state or province,
and a zip or postal code, and sometimes a country.

You also make the point:

Furthermore, the reference value produced by a patterned Address Signal
vector is context-dependent. Much of that context is also shared by the
memory cells of other ECUs, which allows them as a group to request
perceptual values that have in the past been successfully coordinated.

So if I understand you, making the address signal a vector creates a
pattern that may be shared by other contextually relevant patterns within
associative memory. By bringing up that context jointly, it allows for
several perceptual values to be specified and coordinated jointly.

You then seem to emphasize the perceptual contexts that are "temporally
connected," although you note that Bill says nothing in his model requires
the reference signals to be addressed with such temporal connection.

Next, you introduce the interesting observation that when active
controlling is going on in a given environment, the temporal connection
actually exists, because it is provided by the "now" of the present
moment. That raises the possibility that a "now" designation could become
a point of reference for associative memory, almost like an index, for
subsequent perceptual coordination.

You raise as a brief aside something that Bill spends more time on, when
you say that "when we are operating in imagination mode, no such temporal
coordination of the recovery address signals is imposed." That may be why
imagination can easily divert into flights of fancy, because realistic
contextual coordination is absent. By definition, a control loop in
imagination mode is _not_ closed through the environment with its
inconvenient physical properties. Think, and it is done! This may also be
why it can be hard to get the right degree of vividness when daydreaming,
because on one's own, we may almost say a-contextually, one must supply the
requisite detail that the lower levels of perception normally take care
of.

Returning to the main thrust of your argument, you ask what is likely to be
stored in associative memory. You rightly imply that it is unlikely, if
not physically impossible, for everything about an extensive flow of
perceptual experience to get stored. So, what is likely to be stored in
there, you ask. And you give an intriguing answer: "Everywhere in the
system, for the most part, changes are what matter." (Forgive me if I hear
an echo of Bateson's seminal phrase, "differences that make a difference.")

It would seem that the computational task of the brain would be greatly
simplified if we only store a subset of what happens; not all the data, but
only changes in the data. Or more specifically, the data that exist at
significant points of change.

As you say:

What this is leading up to is the suggestion that possibl[y] the memories
that are stored are biased in favour of perceptions at the edge -- at
moments or places of change (spatial or temporal "events" at that
level);...

So then, *change events* become the referent for when (or what) to store in
associative memory. A natural question arises -- which change events? --
and you proceed to suggest an answer in your continuation of the above
sentence:

...perhaps the change is not in the perceptual value being stored, but in
the error signal of the ECU that controls this perception or that.

So then, not just any change event, but a significant *change in error
signal* (either its magnitude or its rate of change) would become the
referent for what constellation of co-occurring perceptions would get
stored in associative memory.

You speak of a sudden losing or regaining control, leading to --

a significant change in either the value or the rate of change of the
error variable in a control system, and I propose that such a change would
be an appropriate moment to store a perception in its pattern context, as
either a situation to be recovered or a situation to be avoided.

The significant change in error becomes in effect a time index for when
to 'take the snapshot' of the flow of current perception, and store it
within associative memory. The pattern of perceptions in that snapshot is
contextually and temporally coordinated in the perceptual "now," to use
your word, so that a fragment of what is recorded (as per Powers' notion of
associative addressing) could become a reference address signal, to summon
up that context again.

Have I got this right? Are these paraphrases close to what you are driving
at in your post?

Assuming for the moment a basic 'yes' answer, the following elaborations or
questions would need to be sorted out:

How much of a change in error signal would it take to be considered
significant? After all, error signals are reducing or increasing all the
time, during every attempt at control. Not everything about every act of
control gets recorded, presumably. So is there a threshold effect, or a
kindling effect of some kind?

Secondly, what could be a mechanism for a change in error to propagate such
contextual storage? Error signals derive from whatever neurological
features instantiate comparators, in Bill's model. Is there something
about the ganglia functioning as the comparators, that have a property for
initiating memory storage, which other summating neurons don't have? My
speculations feel pretty far-fetched here.

At any event, those further questions are not the key ones at this point.
You have presented a very intriguing post, which takes the discussion of
memory storage and comparable reference signals for addressing and
accessing it in some very interesting directions.

All the best,

Erling

[Martin Taylor 2013.11.25.10.00]

[From Erling Jorgensen (2013.11.24 1955EST)]

Martin Taylor 2013.11.17.09.13

Hi Martin,

I was intrigued by your post about a Reference vector, which provides a
context-relevant Address Signal to lower level memory. ...

I'll begin, however, with a rebuttal-type response on this issue of
commensurability, in your reply to my post of [Erling Jorgensen (2013.11.16
1100EST)].

To begin with a comment on what Erling says, I do not see a problem of
incommensurability anywhere in the PCT hierarchy, because all the signals
are either neural firings or variations of potential within a neural cell
body.

Yes, both you and Rick Marken (2013.11.19.0930) hold the view that Powers'
specification of "neural current" creates a common unit among neural
control systems and thus solves the problem of incommensurability.

I get the impression that any difficulty between us is in our understanding of the word "incommensurable". To me, it means being measurable in the same units, such as amps, seconds, kilometers, etc. In that sense, no signals measurable in, for example, "spikes per second" are incommensurable. I think Rick probably had in mind the same meaning of the word.

However, Bill himself seems to have thought there was still "a problem that
has quietly been lurking in the background" (B:CP, 1st ed., p. 217). While
the common currency of neural currents was necessary, it seems not to have
been sufficient, to make sure reference signals are talking with comparable
perceptual signals within Bill's model.

In the quotations from B:CP chapter 15 that I laid out in my previous post
on Memory & Address Signals, I list what Bill seemed to think were
additional constraints on how reference signals were to operate in his
model.

Powers says quite plainly, "Memory, in order to be useful, must be under
control of some orderly addressing mechanism. That is the principle I wish
to communicate" (p. 216, emphasis omitted). So here he is saying that, in
addition to having a comparable current, reference signals must specify a
comparable locale. The address must be right.

Surely all that is the business of evolution and reorganization? Whether we are talking about direct connections of a collection of higher-level outputs into a summed reference value in the manner of Rick's spreadsheet example, or about using those same outputs to address (both for storage and for retrieval) Bill's associative memories, the same arguments apply. What works to control perceptions that help the organism and its descendants to survive will be what survives the (individual and evolutionary) reorganization process.

The "orderly addressing mechanism" is exactly the vector of outputs that selects where the perceptual value for a specific elementary control unit is stored, and whence it will be retrieved at need. I am guessing, but my guess is that Bill may have been misled by his own diagram 15.3, which showed only a single value as the address into the associative memory. He may have missed that his distribution of outputs to the next lower level implied the use of several addressing inputs to the memory unit from the next higher level -- an address vector. That is "orderly" whereas the use of a single (perhaps noisy) value is not.

To me, this has no connection with the word "incommensurable", as it seems to have for you.

So then, in order to avoid a problem of incommensurability between
reference signals and perceptual signals, the following conditions must be
met:

1) Comparable Current

2) Comparable Locale

3) Comparable Pattern (or fragment of a pattern)

It seems, Martin, that you are adding a fourth possible condition:
Comparable Context. And here I'm shifting from rebuttal to review, to
check my understanding of your ideas.

No, I'm not adding any new condition. All three of your criteria are consequences of Bill's proposal when the implied vector addressing is taken into account. "Comparable context" is just another way of saying "same address" (or, see below, close to the same -- same being "Identical context").

···

---------------------

You make the point that a memory address constructed out of a vector of
higher-level outputs is a more stable reference than a single-valued
output. The address becomes much more specific as a result. To note a
trivial analogy, this is why postal addresses don't just have a house (or
apartment) number. That would lead to too many false negatives. The post
office also wants a street name, and a city name, and a state or province,
and a zip or postal code, and sometimes a country.

There is a difference, in that the postal values are inclusion domains rather than a vector of independent criteria.

To make the postal analogy closer, the address might be given as "Number 162, two stories with a gable for an attic window on a third story, a red front door that has a brass knocker, a brick facade over a breeze block foundation, a cedar tree in the front yard, green window trim, and a Porsche likely to be in the driveway." There might be only one such house in the world, or none, or several, but in the general area where the postman works (analogous to, say, auditory or visual memory) the description is likely to exist and to be unique. But there might be no such houses, the owner of the designated house having repainted the window trim red to match the door, or having traded the Porsche for a BMW.

If the mail is to be delivered, retrieval from such an address cannot be required to be exact. The more matches to individual descriptive criteria the better. There may be many houses with red doors, but there are probably not very many with that house number and several of the other descriptive criteria. If the mailman delivers to a two-story house with a gabled attic window, a brick facade over breezeblock, a Chev in the driveway, a blue door with a steel knocker, number 162, one cedar in the front yard, and white window trim, he will probably be delivering to the correct house.

Yes, the mail might be delivered to the wrong address if the postman accepts houses that agree on only six of the ten criteria, or is given only six of them as the address for delivery. But do we not also sometimes retrieve wrong memories that fit much of a retrieval context, but that were actually stored in a slightly different context? In the brain context, the basic address is localized in the sense of city-state-country-postal code, because of the fact that storage probably depends only on those higher-level systems that contribute to retrieval (but again, see below for a caveat).

You also make the point:

Furthermore, the reference value produced by a patterned Address Signal
vector is context-dependent. Much of that context is also shared by the
memory cells of other ECUs, which allows them as a group to request
perceptual values that have in the past been successfully coordinated.

So if I understand you, making the address signal a vector creates a
pattern that may be shared by other contextually relevant patterns within
associative memory. By bringing up that context jointly, it allows for
several perceptual values to be specified and coordinated jointly.

Yes. That seems to be a consequence of the usually accepted structure of the control hierarchy.

You then seem to emphasize the perceptual contexts that are "temporally
connected," although you note that Bill says nothing in his model requires
the reference signals to be addressed with such temporal connection.

If I may make a general comment on Bill's writings, in B:CP and elsewhere, including CSGnet, I have felt that in his need to destroy the notion that the environment determines everything in a stimulus-response way, he ignores the real constraints imposed on control by the environment. By ignoring the fact that what we perceive as the "now" is mostly influenced by the recent history of what we have sensed from the environment, he loses the imposed temporal connections among simultaneous perceptions. However, one should not make the opposite error, either. If my guess about storage being occasioned by sharp changes in error values ("error events") is right, not all elementary control units will incur such events at the same time. Coordinated perceptions and perceptions that support related higher-level controlled perceptions are likely to have error events at closely linked times, but unrelated perceptions may well continue controlling well, without experiencing storage-inducing error events.

The point is that nothing in the model requires different perceptions to be stored at the same moment, but it is rather like the adage "The family that eats together stays together", which may be true more often than chance, but is by no means guaranteed to be true. Perceptions that tend to be disturbed together or whose reference values are likely to be changed together ar perceptions that would be likely to be stored at the same time if the error-event suggestion has any merit.

There is, however, an issue that is not covered at the moment in my suggestions (or in Bill's). The context includes perceptions that are not being controlled at the storage moment, but that are likely to be included in storage and retrieval addressing. Being in a particular place quite often evokes memories of similar places, neither the retrieved memories nor the current perceptions that evoke the memories being controlled. There is more to the story than we have yet discussed. Are currently uncontrolled perceptions that form part of the addressing context themselves stored somewhere, or are they used only for addressing? Is what is stored only what was in consciousness, and is consciousness related to error events? The uncontrolled part of the context must be important, but at the moment it doesn't sit comfortably on the fragile limb on which I depend.

Next, you introduce the interesting observation that when active
controlling is going on in a given environment, the temporal connection
actually exists, because it is provided by the "now" of the present
moment. That raises the possibility that a "now" designation could become
a point of reference for associative memory, almost like an index, for
subsequent perceptual coordination.

Yes, except that for it to be usable as an index, it has to be perceived. Do we perceive a value for "now" that could be used as an index for later retrieval? I doubt it. What I personally perceive consciously is a flow of events, edges in time and space unless I direct my perception to some static aspect of my perceptible environment. Over time, some events perceptually precede others, so I perceive "before" and "after", but I do not seem to perceive a "when", other than by proxy if I see a clock or hear a time announcement. But I do seem to have some rudimentary and occasional perception of "when", as evidenced by a December day in Copenhagen, which is much further north than Toronto. I was downtown around noon, but I felt it was about 4 pm (16:00), presumably because of the quality of the sunlight. And I do seem to have some sense of how long a second feels, but not a minute or an hour, those depending so much on what is happening and what I am doing.

On the other hand, the set of perceptions that are part of the conscious "now" could be part of the storage address, or "index" without being identified as being at any specific "when" (See the comment above about an issue not covered).

You raise as a brief aside something that Bill spends more time on, when
you say that "when we are operating in imagination mode, no such temporal
coordination of the recovery address signals is imposed."

Yes. I had just wanted to leave a pointer to what Bill said on that topic.

Returning to the main thrust of your argument, you ask what is likely to be
stored in associative memory. You rightly imply that it is unlikely, if
not physically impossible, for everything about an extensive flow of
perceptual experience to get stored. So, what is likely to be stored in
there, you ask. And you give an intriguing answer: "Everywhere in the
system, for the most part, changes are what matter." (Forgive me if I hear
an echo of Bateson's seminal phrase, "differences that make a difference.")

I suppose instead of "are what matter" I should have said "are likely to affect the downstream neural signals", but that's a rather more long-winded way of saying the same thing!

As you say:

What this is leading up to is the suggestion that possibl[y] the memories
that are stored are biased in favour of perceptions at the edge -- at
moments or places of change (spatial or temporal "events" at that
level);...

So then, *change events* become the referent for when (or what) to store in
associative memory. A natural question arises -- which change events? --
and you proceed to suggest an answer in your continuation of the above
sentence:

...perhaps the change is not in the perceptual value being stored, but in
the error signal of the ECU that controls this perception or that.

So then, not just any change event, but a significant *change in error
signal* (either its magnitude or its rate of change) would become the
referent for what constellation of co-occurring perceptions would get
stored in associative memory.

That's the notion, but "constellation of co-occurring perceptions" is a consequence, not a presupposition. Each "memory cell" is a part of one elementary control unit, just as is the perceptual function, the comparator, or the output function. It is one possible type of what we called, years ago, the "Reference Input Function" (RIF). One would expect that there are other types, such as summing the convergent outputs of higher-level units, or possibly correlators that produce an output proportional to the multiple correlation of its inputs. Which, if any, of these types of RIF actually exist in the brain remains to be determined. But for now I am trying to work out some implications of Bill's associative memory proposal, ignoring other types of RIF. I could well imagine that associative memories are the RIF type only at higher hierarchic levels if they exist at all, and that lower levels have different analogue functions, perhaps different for different hierarchic levels.

As an aside, I suspect it will prove as hard to specify exactly how RIFs operate as it is to specify how PIFs operate. We might just have to handwave and say "but that's the result of the function" in the same way we do for, say, configuration perception. Some day we may know how it works, but for now all we know is that we can perceive configurations.

The "constellation" is just another word for a group of memory cells that happen to get something stored in them (or retrieved from them) more or less simultaneously.

You speak of a sudden losing or regaining control, leading to --

a significant change in either the value or the rate of change of the
error variable in a control system, and I propose that such a change would
be an appropriate moment to store a perception in its pattern context, as
either a situation to be recovered or a situation to be avoided.

The significant change in error becomes in effect a time index for when
to 'take the snapshot' of the flow of current perception, and store it
within associative memory. The pattern of perceptions in that snapshot is
contextually and temporally coordinated in the perceptual "now," to use
your word, so that a fragment of what is recorded (as per Powers' notion of
associative addressing) could become a reference address signal, to summon
up that context again.

Have I got this right? Are these paraphrases close to what you are driving
at in your post?

Pretty well. Reading between the lines of your paraphrase, I don't get the sense that you got the point that the perceptual value stored in the memory cell of a control unit is the value of the perception being controlled by that unit, as in Bill's diagram. The vector ("constellation") of coordinated control is a consequence of the fact that something happened to all those control units at more or less the same time. On other occasions, any particular memory cell might belong to a different "constellation".

Assuming for the moment a basic 'yes' answer, the following elaborations or
questions would need to be sorted out:

How much of a change in error signal would it take to be considered
significant? After all, error signals are reducing or increasing all the
time, during every attempt at control. Not everything about every act of
control gets recorded, presumably. So is there a threshold effect, or a
kindling effect of some kind?

Good question. Since the notion of error events being the trigger for a storage event is itself highly speculative, any answer must be even more speculative -- the limb I am out on gets ever longer and weaker.

But here's a kind of suggestion in place of an answer. When we talk about neural currents, we are actually using one number to represent the firing of many neurons that have similar, but never identical, functions (i.e. connections and chemistry). Likewise, a memory "cell" is a shorthand for changes in the synapses (presumably) of many neurons. So, when we talk about storage of a perception, we are talking about a property combined over many locations. Hence it is unlikely to show a threshold. More probably, a "perception" is stored strongly when very many neurons are involved because there are big changes in the error signal, and stored weakly and reversibly when the error signal changes are small.

Neurons themselves have firing thresholds, so one might imagine that there could be a low threshold of error change below which storage is essentially zero. Because of adaptation, that threshold would probably depend on the usual variability of that particular error signal. Poorly controlled perceptions would have greater error variability than well controlled ones, and would be expected to have their perceptual value stored only when changes are greater, either by regaining good control or by larger than usual abrupt changes in reference or disturbance.

Secondly, what could be a mechanism for a change in error to propagate such
contextual storage? Error signals derive from whatever neurological
features instantiate comparators, in Bill's model. Is there something
about the ganglia functioning as the comparators, that have a property for
initiating memory storage, which other summating neurons don't have? My
speculations feel pretty far-fetched here.

Again, we are in a highly speculative realm, but without going into magical thinking, I just suggest that connection patterns are implicit in the construction of all control systems. The perceptual signal is constructed by the operation of neurons connected in specific ways, and the perceptual signal together with the reference signal connects to something that acts as a comparator only because that's the way the connections have grown (evolved, reorganized). I see no difference in principle between those connections and a putative connection to the memory cell from the perceptual signal and the error signal.

At any event, those further questions are not the key ones at this point.
You have presented a very intriguing post, which takes the discussion of
memory storage and comparable reference signals for addressing and
accessing it in some very interesting directions.

All the best,

Erling

Thanks for the comments. I hope I have somewhat addressed your main issues, and that the following messages on the vector-addressed memory theme (emotions already sent, and avoidance control under construction) also will be found intriguing.

Martin

[From Erling Jorgensen (2013.11.27 1135EST)]

Martin Taylor 2013.11.25.10.00]

[From Erling Jorgensen (2013.11.24 1955EST)]

Hi Martin,

Yes, both you and Rick Marken (2013.11.19.0930) hold the view that Powers'
specification of "neural current" creates a common unit among neural
control systems and thus solves the problem of incommensurability.

I get the impression that any difficulty between us is in our
understanding of the word "incommensurable".

I am reminded of the line in the semi-classic movie, "The Princess Bride,"
where one of the characters keeps saying about all kinds of obstacles that
keep coming up, "It's inconceivable!!" And finally one of the other
characters says, "I don't think that word means what you think it does."

Perhaps it is a matter of terminology. Though I must confess, I am hard
pressed to find a better word to capture this issue of having the right
reference signal being compared to the right perceptual signal. It goes
beyond having "neural current" as a common currency. The definition of
incommensurable that I'm thinking of is "impossible to measure _or
compare_" (emphasis added). If a signal goes to the wrong address, it does
not get compared in the right way for controling the perception associated
with that ECU (Elementary Control Unit).

Powers says quite plainly, "Memory, in order to be useful, must be under
control of some orderly addressing mechanism. That is the principle I wish
to communicate" (p. 216, emphasis omitted). So here he is saying that, in
addition to having a comparable current, reference signals must specify a
comparable locale. The address must be right.

Surely all that is the business of evolution and reorganization?

Forgive me, Martin, but that sounds like handwaving: 'And so, we see the
signal arrives just where it must in order to function properly.'

I like what you are saying about a vector-based reference signal, and about
its role in associative memory. But I don't think that discounts what I am
claiming to be Bill's conditions for such reference signals, that they must
involve a comparable current, a comparable locale, and a comparable
pattern.

No, I'm not adding any new condition. All three of your criteria are
consequences of Bill's proposal when the implied vector addressing is
taken into account. "Comparable context" is just another way of
saying "same address" ...

I don't think I agree that "comparable context" means "same (or close to
the same) address." The temporal context in memory of a symphony may be a
particular concert hall, and a special evening out, and being intrigued by
the conductor's movements. But none of those seem to be the "same address"
for recognizing phrases of the music from memory. They may be the
surrounding context of related perceptions, sort of like being 'in the
neighborhood,' but they are not the memory address for auditory
perceptions. Or so it seems to me.

There is a difference, in that the postal values are inclusion domains
rather than a vector of independent criteria.

I'm not sure I understand the difference here, although I don't think our
conceptions are all that disparate. Your extension of the postal analogy
does not drive it home for me. Are you saying that "inclusion domains" all
must be true, whereas with "a vector of independent criteria" it may be
sufficient for enough of them to overlap? If so, I can accept that
distinction for vector-based reference signals.

If my guess about storage being occasioned by sharp changes in error
values ("error events") is right, not all elementary control units will
incur such events at the same time.

I like your term "error events," although I am still trying to wrap my
thinking around "sharp changes in error values." This continues my query
about what is a significant enough change in the error. I'm trying to
figure out if we could substitute the term "unexpectedly large error
events," but I don't like the implications of going that route. It would
seem to imply that the size of the error itself becomes a controlled
variable, relative to what is 'expected,' and I can imagine all kinds of
instabilities being introduced into the loop by such a route. I don't
think you are saying that. But I'm not sure what constitutes a "sharp
change."

It gets me closer to your way of thinking to see the phrase "storage-
inducing error events."

That raises the possibility that a "now" designation could become a point
of reference for associative memory, almost like an index, for subsequent
perceptual coordination.

Yes, except that for it to be usable as an index, it has to be perceived.

I don't think I was understanding a "now" designation as itself an index.
I was following the flow of your argument: a) Temporally connected address
signals could exist if there is a common referent; for instance (as a first
approximation), the now of the present moment. b) What if (as a second
approximation) the common temporal referent were a particular change
event? c) What if (as a third approximation) it were specifically a
significant change in error signal, or as you call it in this post,
an "error event"? d) Any of these, but especially the last of the three,
would be a significant temporal point for contextually coordinated storage
in memory. That's the sense in which I meant a time index -- as I said
later in my post:

The significant change in error becomes in effect a time index for when
to 'take the snapshot' of the flow of current perception, and store it
within associative memory.

Do we perceive a value for "now" that could be used as an index for later
retrieval? I doubt it.

I agree; usually not.

So then, not just any change event, but a significant *change in error
signal* (either its magnitude or its rate of change) would become the
referent for what constellation of co-occurring perceptions would get
stored in associative memory.

That's the notion, but "constellation of co-occurring perceptions" is a
consequence, not a presupposition.

Yes. That's why I used to passive voice -- "...would get stored..."

Each "memory cell" is a part of one elementary control unit, just as is
the perceptual function, the comparator, or the output function. It is one
possible type of what we called, years ago, the "Reference Input Function"
(RIF).

This is a helpful description for me seeing what you are getting at.

As an aside, I suspect it will prove as hard to specify exactly how RIFs
operate as it is to specify how PIFs operate.

Oh, aren't you a joy! You're probably right about that, unfortunately. We
need someone like Jeff Hawkins or his team, but looking for physiological
analogues of the PCT concepts and 'convenient fictions.'

The "constellation" is just another word for a group of memory cells that
happen to get something stored in them (or retrieved from them) more or
less simultaneously.

Again, this is helpful.

How much of a change in error signal would it take to be considered
significant? ...

Good question. Since the notion of error events being the trigger for a
storage event is itself highly speculative, any answer must be even more
speculative -- the limb I am out on gets ever longer and weaker.

LOL! Well, my friend the squirrel, you seem to be able to dance pretty
well out there!

Secondly, what could be a mechanism for a change in error to propagate
such contextual storage? Error signals derive from whatever neurological
features instantiate comparators, in Bill's model. ...

Again, we are in a highly speculative realm, but without going into
magical thinking, I just suggest ...[snip]... a putative connection to the
memory cell from the perceptual signal and the error signal.

This is a hint at the wiring that I had not noticed before: Memory cells
for storing particular perceptual values may be created out of joint
connections from the PIF (perceptual input function) and the Comparator of
a particular Elementary Control Unit. There may be something here.

Thanks for the comments. I hope I have somewhat addressed your main
issues, and that the following messages on the vector-addressed memory
theme (emotions already sent, and avoidance control under construction)
also will be found intriguing.

...Always good doing business with you, Martin.

All the best,
Erling

[Martin Taylor 2013.11.2714.22]

[From Erling Jorgensen (2013.11.27 1135EST)]
Martin Taylor 2013.11.25.10.00]
[From Erling Jorgensen (2013.11.24 1955EST)]
Hi Martin,
Yes, both you and Rick Marken (2013.11.19.0930) hold the view that Powers'
specification of "neural current" creates a common unit among neural control systems and thus solves the problem of incommensurability.
I get the impression that any difficulty between us is in our understanding of the word "incommensurable".
Perhaps it is a matter of terminology. ... If a signal goes to the wrong address, it does not get compared in the right way for controling the perception associated with that ECU (Elementary Control Unit).
You seem to use the word "incommensurable" where I might substitute

“not useful”.

I imagine signals going to a useless address happens a lot in the

early wiring of a brain. What would be the result? Actions that
don’t reliably influence whatever perception is wired to the
comparator with that reference wiring. What does “standard”
reorganization theory say would happen then? As I understand it, it
says that these connections would disappear randomly over time,
whereas connections that did result in actions that reliably
influenced the perceptions would tend to stick around. Is it a
coincidence that there seems to be more pruning than augmentation of
the connections in the infant brain?


Powers says quite plainly, "Memory, in order to be useful, must be under
control of some orderly addressing mechanism. That is the principle I wish
to communicate" (p. 216, emphasis omitted). So here he is saying that, in
addition to having a comparable current, reference signals must specify a
comparable locale. The address must be right.
Surely all that is the business of evolution and reorganization?
Forgive me, Martin, but that sounds like handwaving: 'And so, we see the signal arrives just where it must in order to function properly.'
If you want to start a thread on reorganization, I would happy to

participate because I do have some ideas on it, but I thought that
Bill had demonstrated pretty well that in a toy system of a few tens
of randomly connected control units, e-coli reorganization before
long resulted in a structure in which not only were the perceptions
and references appropriately linked through comparators to actuators
that influenced the right perceptions, but also that the reorganized
control units tended to operate orthogonally.

I didn't think I was invoking magic so much as the nearly inevitable

consequences of a “keep what works, discard what doesn’t” process
that works over both evolutionary and individual time.


No, I'm not adding any new condition. All three of your criteria are consequences of Bill's proposal when the implied vector addressing is taken into account. "Comparable context" is just another way of saying "same address" ...
I don't think I agree that "comparable context" means "same (or close to the same) address." The temporal context in memory of a symphony may be a particular concert hall, and a special evening out, and being intrigued by the conductor's movements. But none of those seem to be the "same address" for recognizing phrases of the music from memory.
Isn't it true that if you imagine the symphony you are likely to

remember that environment, and vice versa? For myself, I vividly
remember the first concert my father took me to, in 1947. We sat
right behind the French Horn in Edinburgh’s Usher Hall (not a name I
could have called up before remembering the event). One of the
pieces they played was Ibert’s Divertissements, in which the horn
player stands up and blows a railway whistle as loud as he can. I
cannot visualize the scene without hearing the music, and if I hear
that music on the radio or imagine it in my head as I am doing at
the moment, I see the scene.

However, you do touch on another issue, which is the part played by

uncontrolled perceptions in determining the address for storage and
recovery. They do form part of the perceptual context, but they are
not represented among the inputs to the Reference Input Function of
any control unit. I think they should form part of the addressing
context, but where would the line be drawn between what is and is
not included? The same point applies to the inclusion of the
associated emotional perception in the address [Martin Taylor
2013.11.21.14.47].

They may be the surrounding context of related perceptions, sort of like being 'in the neighborhood,' but they are not the memory address for auditory perceptions. Or so it seems to me.
Why specifically not for auditory perceptions? What is special about

them?


There is a difference, in that the postal values are inclusion domains rather than a vector of independent criteria.
I'm not sure I understand the difference here, although I don't think our conceptions are all that disparate. Your extension of the postal analogy does not drive it home for me. Are you saying that "inclusion domains" all must be true, whereas with "a vector of independent criteria" it may be sufficient for enough of them to overlap? If so, I can accept that distinction for vector-based reference signals.
In my revised postal analogy, none of the criteria overlap, whereas

in the city-state-country set of criteria, they MUST overlap. The
successive levels of the address are refinements of one criterion
(location) rather than independent criteria. If the address
specifies a postal code that isn’t in the specified city, and the
city isn’t in the specified state, what can the postman do? The
address is simply wrong, and if it is fixable, the fix must come
from somewhere else – for example, the postman might note that
Tronto is a plausible misspelling of Toronto, and guess that Toronto
was intended by the sender. With a bunch of orthogonal criteria, you
only need enough matches to reduce the number of reasonable
possibilities to one.


If my guess about storage being occasioned by sharp changes in error values ("error events") is right, not all elementary control units will incur such events at the same time.
I like your term "error events," although I am still trying to wrap my thinking around "sharp changes in error values." This continues my query about what is a significant enough change in the error.
Remember that neural currents are a shorthand way of talking about a

lot of firings of individual neurons that do more or less the same
thing. If you don’t think of the error value as a neural current,
but instead think of many individual neurons that fire at their own
rates according to the errors in their individual micro-control
units, you can see that “significant change” just means that a lot
of them change at the same time, and insignificant change means only
a few of them do.

As for what detects changes in firing rates for the individual

control units that together make the box in a diagram, I think
trying to answer that question is somewhere out on a twig of the
branch I am hoping does not break under the weight of my
suggestions.

I'm trying to figure out if we could substitute the term "unexpectedly large error events," but I don't like the implications of going that route. It would seem to imply that the size of the error itself becomes a controlled variable, relative to what is 'expected,' and I can imagine all kinds of instabilities being introduced into the loop by such a route. I don't think you are saying that.
You are right, I'm not. Let's see whether I can reach out to that

twig without having the branch break…

One of the starting points of my branch-climbing adventure was the

concept that most neurons have a firing rate that does not depend so
much on the intensity of their input, but on changes in that
intensity. If the input increases, the output firing rate increases
and then subsides; if the input decreases the output drops and then
recovers, perhaps not to the same resting level, but nevertheless
subsidence or recovery after a change does seem to occur in at least
some neurons. If we assume that some such time course is normal for
the neurons that output the error signal, we can then imagine the
existence of a receiving neuron that fires only if its input is
sufficient, and that the firing of that neuron is the
“storage-inducing event”. The input to such a neuron might be from
the neuron earlier mentioned and additionally from neuron that is
inhibited by the output of the first neuron. The loss of that
inhibition might also induce a storage event. It might occur for
some of the control units but not for others in the group that
together define the neural currents.

Here's a figure to make the suggestion easier to interpret. The tiny

circles mean inhibition. If the “comparator” neuron is in a steady
state, I’m assuming that it does not fire enough to make the storage
signal neuron fire. But if its rate increases or decreases, the
input to the storage signal neuron increases, assuming reasonable
non-linearities in the response of the inhibited intermediary
neuron. I’ve no idea whether such circuits or ones that would do the
same job better actually exist, but I suppose they could. And there
would probably be hundreds or thousands of inputs to each of these
neurons, so in no way would it be as simple as I have drawn it.

![StorageSignal.jpg|428x141](upload://5mnoXeKwRbaNzAuLVBVOZxDJ2lr.jpeg)

Now that's hand-waving with white gloves on! It takes real magic to

prevent the twig from breaking under the weight of all that.


...Always good doing business with you, Martin. All the best, Erling
I wonder if any of this bears any relation to the truth? So far as I

can see, it’s not inconsistent with anything I know, but then I
don’t know as much as I should in order to test its plausibility
very well. So, don’t look behind the curtain – that’s what
hand-waving is supposed to discourage.

Martin

[From Rick Marken (2013.11.27.2130)]

I think the elephant in the room when it comes to PCT and memory is that there is really no PCT model of how memory works. This can be seen by considering Fig. 15.3 in B:CP. This diagram shows how storage and retrieval fit into the PCT model, in terms of signal pathways. But it doesn’t show how these functions are carried out. For example, what throws the memory switch from reference signal to perceptual signal? And why? What switches the perceptual switch from perceptual signal to retrieved reference signal? And why? And what about the storage signal; there is no switch shown, implying that perceptions are always being stored in memory. Is this really the case? If so, then what’s going on when I actively try to remember a phone number, repeating it to myself in my head?

I think what is needed is a model of the system – probably a control system – that is responsible for the operation of the switches and the storage process shown in Fig 15.3. I think this should be a control system based on my subjective experience of purposefully trying to remember things (controlling for being able to recall the phone number, for example); I am controlling for being able to recall things at a later time. The recall process also is purposeful; I am controlling for retrieving the correct number. How is that done? Imagining is also a purposeful process; I control for going into imagination mode (controlling for perceptions in my head without actually controlling sensory based perceptions). When people say “Let me think about that” I imagine they are doing what I do in such circumstances; purposefully switch to imagination mode. HOw is this done?

I think the lack of a memory model of the kind I describe above – a model that controls using the “switches” in Figure 15.3 as the means of control-- is the main reason why there has been no PCT based research on memory; there really is no model of memory to test. The model in Fig 15.3 accounts for the fact that we have memory (the storage and retrieval connections), that we can imagine, and so on. But it seems to me that the PCT model of memory – as shown in Figure 15.3 – is missing a model of the system that allows people to purposefully memorize and recall (imagine) perceptions.

I think the development of a testable model of purposeful memorization would be a very productive way to expand the PCT model. Of course, I may be wrong; maybe there is a model of memory already existing in PCT. But if there is, it should be able to explain and predict the results of some experiments on memory. One of the problems I’ve had with the current discussion of addressing systems is that it was not clear to me how one would test the different proposals. I think if a true working model of memorization/recall were developed it would have behavioral implications that could be tested; and the model might incorporate one or another of these addressing schemes, which might also have different behavioral implications.

So I encourage the development of a PCT-based memory model that has testable behavioral implications, serving as the basis for a PCT- based research program on memory.

Best regards

Rick

···

On Wed, Nov 27, 2013 at 10:00 AM, Erling Jorgensen ejorgensen@riverbendcmhc.org wrote:

[From Erling Jorgensen (2013.11.27 1135EST)]

Martin Taylor 2013.11.25.10.00]

[From Erling Jorgensen (2013.11.24 1955EST)]

Hi Martin,

Yes, both you and Rick Marken (2013.11.19.0930) hold the view that Powers’

specification of “neural current” creates a common unit among neural

control systems and thus solves the problem of incommensurability.

I get the impression that any difficulty between us is in our

understanding of the word “incommensurable”.

I am reminded of the line in the semi-classic movie, “The Princess Bride,”

where one of the characters keeps saying about all kinds of obstacles that

keep coming up, “It’s inconceivable!!” And finally one of the other

characters says, “I don’t think that word means what you think it does.”

Perhaps it is a matter of terminology. Though I must confess, I am hard

pressed to find a better word to capture this issue of having the right

reference signal being compared to the right perceptual signal. It goes

beyond having “neural current” as a common currency. The definition of

incommensurable that I’m thinking of is "impossible to measure _or

compare_" (emphasis added). If a signal goes to the wrong address, it does

not get compared in the right way for controling the perception associated

with that ECU (Elementary Control Unit).

Powers says quite plainly, "Memory, in order to be useful, must be under

control of some orderly addressing mechanism. That is the principle I wish

to communicate" (p. 216, emphasis omitted). So here he is saying that, in

addition to having a comparable current, reference signals must specify a

comparable locale. The address must be right.

Surely all that is the business of evolution and reorganization?

Forgive me, Martin, but that sounds like handwaving: 'And so, we see the

signal arrives just where it must in order to function properly.’

I like what you are saying about a vector-based reference signal, and about

its role in associative memory. But I don’t think that discounts what I am

claiming to be Bill’s conditions for such reference signals, that they must

involve a comparable current, a comparable locale, and a comparable

pattern.

No, I’m not adding any new condition. All three of your criteria are

consequences of Bill’s proposal when the implied vector addressing is

taken into account. “Comparable context” is just another way of

saying “same address” …

I don’t think I agree that “comparable context” means "same (or close to

the same) address." The temporal context in memory of a symphony may be a

particular concert hall, and a special evening out, and being intrigued by

the conductor’s movements. But none of those seem to be the “same address”

for recognizing phrases of the music from memory. They may be the

surrounding context of related perceptions, sort of like being 'in the

neighborhood,’ but they are not the memory address for auditory

perceptions. Or so it seems to me.

There is a difference, in that the postal values are inclusion domains

rather than a vector of independent criteria.

I’m not sure I understand the difference here, although I don’t think our

conceptions are all that disparate. Your extension of the postal analogy

does not drive it home for me. Are you saying that “inclusion domains” all

must be true, whereas with “a vector of independent criteria” it may be

sufficient for enough of them to overlap? If so, I can accept that

distinction for vector-based reference signals.

If my guess about storage being occasioned by sharp changes in error

values (“error events”) is right, not all elementary control units will

incur such events at the same time.

I like your term “error events,” although I am still trying to wrap my

thinking around “sharp changes in error values.” This continues my query

about what is a significant enough change in the error. I’m trying to

figure out if we could substitute the term "unexpectedly large error

events," but I don’t like the implications of going that route. It would

seem to imply that the size of the error itself becomes a controlled

variable, relative to what is ‘expected,’ and I can imagine all kinds of

instabilities being introduced into the loop by such a route. I don’t

think you are saying that. But I’m not sure what constitutes a "sharp

change."

It gets me closer to your way of thinking to see the phrase "storage-

inducing error events."

That raises the possibility that a “now” designation could become a point

of reference for associative memory, almost like an index, for subsequent

perceptual coordination.

Yes, except that for it to be usable as an index, it has to be perceived.

I don’t think I was understanding a “now” designation as itself an index.

I was following the flow of your argument: a) Temporally connected address

signals could exist if there is a common referent; for instance (as a first

approximation), the now of the present moment. b) What if (as a second

approximation) the common temporal referent were a particular change

event? c) What if (as a third approximation) it were specifically a

significant change in error signal, or as you call it in this post,

an “error event”? d) Any of these, but especially the last of the three,

would be a significant temporal point for contextually coordinated storage

in memory. That’s the sense in which I meant a time index – as I said

later in my post:

The significant change in error becomes in effect a time index for when

to ‘take the snapshot’ of the flow of current perception, and store it

within associative memory.

Do we perceive a value for “now” that could be used as an index for later

retrieval? I doubt it.

I agree; usually not.

So then, not just any change event, but a significant *change in error

signal* (either its magnitude or its rate of change) would become the

referent for what constellation of co-occurring perceptions would get

stored in associative memory.

That’s the notion, but “constellation of co-occurring perceptions” is a

consequence, not a presupposition.

Yes. That’s why I used to passive voice – “…would get stored…”

Each “memory cell” is a part of one elementary control unit, just as is

the perceptual function, the comparator, or the output function. It is one

possible type of what we called, years ago, the “Reference Input Function”

(RIF).

This is a helpful description for me seeing what you are getting at.

As an aside, I suspect it will prove as hard to specify exactly how RIFs

operate as it is to specify how PIFs operate.

Oh, aren’t you a joy! You’re probably right about that, unfortunately. We

need someone like Jeff Hawkins or his team, but looking for physiological

analogues of the PCT concepts and ‘convenient fictions.’

The “constellation” is just another word for a group of memory cells that

happen to get something stored in them (or retrieved from them) more or

less simultaneously.

Again, this is helpful.

How much of a change in error signal would it take to be considered

significant? …

Good question. Since the notion of error events being the trigger for a

storage event is itself highly speculative, any answer must be even more

speculative – the limb I am out on gets ever longer and weaker.

LOL! Well, my friend the squirrel, you seem to be able to dance pretty

well out there!

Secondly, what could be a mechanism for a change in error to propagate

such contextual storage? Error signals derive from whatever neurological

features instantiate comparators, in Bill’s model. …

Again, we are in a highly speculative realm, but without going into

magical thinking, I just suggest …[snip]… a putative connection to the
memory cell from the perceptual signal and the error signal.

This is a hint at the wiring that I had not noticed before: Memory cells

for storing particular perceptual values may be created out of joint

connections from the PIF (perceptual input function) and the Comparator of

a particular Elementary Control Unit. There may be something here.

Thanks for the comments. I hope I have somewhat addressed your main

issues, and that the following messages on the vector-addressed memory

theme (emotions already sent, and avoidance control under construction)

also will be found intriguing.

…Always good doing business with you, Martin.

All the best,

Erling


Richard S. Marken PhD
www.mindreadings.com
The only thing that will redeem mankind is cooperation. – Bertrand Russell

[Martin Taylor 2013.11.28.10.58]

[From Rick Marken (2013.11.27.2130)]

I think the elephant in the room when it comes to PCT and memory is that there is really no PCT model of how memory works.

I whole heartedly agree with you. It is why I have described my musings as "going out on a limb", which in less metaphoric language means "making plausible suggestions without supporting evidence".

This can be seen by considering Fig. 15.3 in B:CP. This diagram shows how storage and retrieval fit into the PCT model, in terms of signal pathways. But it doesn't show how these functions are carried out. For example, what throws the memory switch from reference signal to perceptual signal? And why? What switches the perceptual switch from perceptual signal to retrieved reference signal? And why? And what about the storage signal; there is no switch shown, implying that perceptions are always being stored in memory. Is this really the case? If so, then what's going on when I actively try to remember a phone number, repeating it to myself in my head?
I think what is needed is a model of the system -- probably a control system -- that is responsible for the operation of the switches and the storage process shown in Fig 15.3.

Yes, and also to examine whether such switches exist. The whole structure of Fig 15.3 is a "plausible suggestion" that needs to be tested. .

I think this should be a control system based on my subjective experience of purposefully trying to remember things (controlling for being able to recall the phone number, for example); I am controlling for being able to recall things at a later time.

Yes. What perception are you controlling when you are trying to do this, and what happens when that error is eliminated? How would you disturb it? For me, there is no feeling of "now I remember it", which is the kind of feeling of success I get when I solve a difficult puzzle or execute some fancy passage on the piano. The best I get is "I think now I ought to remember it when I want to recall it". Only later do I know whether the storage control system worked, and even then, failure may mean that the retrieval control system failed when the storage system worked.

The recall process also is purposeful; I am controlling for retrieving the correct number. How is that done? Imagining is also a purposeful process; I control for going into imagination mode (controlling for perceptions in my head without actually controlling sensory based perceptions). When people say "Let me think about that" I imagine they are doing what I do in such circumstances; purposefully switch to imagination mode. How is this done?

All great questions. And these are questions that need to be asked not only in humans but also in non-human species such as crows, dogs, and fish. Crows, for example, seem to be able to discriminate "bad" people from "good" people and to be able to imagine at least some multi-step puzzle solutions. Homing pigeons seem to use landmarks when they are near home (which could be a result of reorganization rather than of episodic memory). Dogs can do some things chimps can't, and at least some fish seem to have at least some episodic memory.

I think the lack of a memory model of the kind I describe above -- a model that controls using the "switches" in Figure 15.3 as the means of control-- is the main reason why there has been no PCT based research on memory; there really is no model of memory to test. ...
I think the development of a testable model of purposeful memorization would be a very productive way to expand the PCT model.

So I encourage the development of a PCT-based memory model that has testable behavioral implications, serving as the basis for a PCT- based research program on memory.

Do you have any ideas as to how to start? Since there exists a reorganization theory with worked examples that seems to cover procedural memory, I guess we should address working and episodic memories, which seem to be involved in Fig 15.3. I know you like to discard everything done by conventional psychologists, but these do seem to be real distinctions, and have a plausible mapping onto Bill's structure.

Martin

[From Rick Marken (2013.11.29.1230)]

Martin Taylor (2013.11.28.10.58)--

RM: I think the elephant in the room when it comes to PCT and memory is that there is really no PCT model of how memory works.

MT: I whole heartedly agree with you. It is why I have described my musings as "going out on a limb", which in less metaphoric language means "making plausible suggestions without supporting evidence".

RM: That's very reassuring. Thanks Martin! And the rest of your post is very useful and consistent with what I was thinking. Again, thanks!

...

RM: So I encourage the development of a PCT-based memory model that has testable behavioral implications, serving as the basis for a PCT- based research program on memory.

MT: Do you have any ideas as to how to start?

RM: Actually, no. I think Warren Mansell has some people working on this at Manchester. All I would suggest is "keep it simple" which means make as few changes as possible to the existing model. Maybe try applying the model to a memory for cursor versus handle movement task. I guess I would indeed look over some of the memory literature to see if there is a hint in there about how the storage and retrieval processes might be controlled.

I'll put this away in the back of my mind and see if my reorganization system come up with anything.

Best

Rick

···

Since there exists a reorganization theory with worked examples that seems to cover procedural memory, I guess we should address working and episodic memories, which seem to be involved in Fig 15.3. I know you like to discard everything done by conventional psychologists, but these do seem to be real distinctions, and have a plausible mapping onto Bill's structure.

--
Richard S. Marken PhD
<http://www.mindreadings.com>www.mindreadings.com
The only thing that will redeem mankind is cooperation.
-- Bertrand Russell

[From Rick Marken (2013.11.19.0930)]

···

B:CP Chapter 15: Memory & Address Signals

Erling Jorgensen (2013.11.16 1100EST)

I’m using the first edition of Behavior: The Control of Perception (1973).

On p. 217, Bill says, almost in passing:

“This, incidentally, solves the problem of translating an error in a higher-order variable into a specific value of a lower-order perception, a problem that has quietly been lurking in the background.”

I’d like to unpack what I think is going on in this passage and in the surrounding pages of B:CP.

First, what problem is being solved? And why is it a problem in the first place? To give it a name, it is the issue of commensurability. Wikipedia defines commensurability as follows: “Two concepts or things are commensurable if they are measurable or comparable by a common standard.”

It’s the problem of interface. How will two things communicate in a way that is understandable to both sides? What language or units or frames do they hold in common?

At this juncture in the history of Perceptual Control Theory, this may seem a trivial problem. After all, we have already had numerous rigorous demonstrations of principle, in the various computer simulations that show just how well a negative-feedback control architecture can operate. There are even simulations of hierarchical arrangements of control loops, where different classes of perception are being defined by the perceptual input functions, but stable control is achieved nonetheless as higher levels set reference standards for lower level loops.

In all these simulations, commensurability between the different levels is presupposed. It’s handled by the computer. The equations assign numbers to the different variables, and the computations proceed straightaway. It is why we operate with an assumption that variables perform as scalar quantities, even if they may represent a more complex form of perception.

RM: I don’t see the problem here. In the simulations the numbers in the control system part of the model represent neural current rates. A reference signal is, thus, the same as a perceptual (and an error) signal. Perceptual and reference signals are “commensurable” because they are the same thing; a neural current. This is not a presupposition of the computer simulation; it’s a basic assumption of the PCT model. It’s true that some neural currents represent more complex perceptions than others; but the complexity of a perception is determined (according to PCT) by the nature of the perceptual function, not by the nature of the neural signal. So a reference signal of 10 (10 pulses perception) is a specification for a perceptual signal of 10, no matter where in the hierarchy this synapse occurs. If this reference signal is the input to a control system whose perceptual signal is the output of a perceptual function at the sensation level – say the sensation of pressure – then the reference signal is a specification for 10 units of pressure signal. If the reference signal is the input to a control system whose perceptual signal is the output of a perceptual function at the principle level – say the principle “honesty is the best policy”-- then the reference signal is a specification for 10 units of honesty signal.

I think Bill went to considerable effort to make sure there was no commensurability problem in PCT. And I don’t think the “reference signal” as address would solve it anyway, if such a problem existed; after all, all that addressing does is point to another neural signal; so you’ve still got neural signals specifying the values of other neural signals.

I actually don’t know what “problem” conceiving reference signals as address solves. I think it’s just an elegant way of incorporating memory into the model.

But I think a point you make in a later post of yours is a very good one; in your post [Erling Jorgensen (2013.11.17 2150EST)] you say:

EJ: One of the large loose ends in PCT & its simulations, to my mind, has been the paucity of realistic input functions.

RM: This is a very good point! I think it could be the basis for a whole new line of research. By realistic, I presume you mean simulations of perceptual functions that do the computations the way a neural network could do it, using only the known properties of neurons to do the calculation. You suggest that we only know how to build realistic perceptual functions like this up to the sensation level. But I think there are good neural models of configuration detectors (perceptual functions). There are probably also good models of transition and sequence detectors (in the speech recognition area for sure). There might even be program detectors (in the virus detection area); and there might even be higher order perceptual functions, like mood detectors, again in the speech recognition area. But now that I think of it, these more advanced perceptual functions are probably not implemented in anything like the way neural nets do it.

Anyway, I think building models with realistic perceptual functions might be a great way for PCT research to interface with other areas of research in robotics, speech recognition, etc. So thanks, Erling, for making this excellent observation.

Best regards

Rick