[Bruce Nevin (2019011.14:40 ET)]
RM: If you get a chance could you give a brief summary of what they had the animals doing. And why you think the animals weren’t doing what the researches thought they were doing.
Irrelevant. The substance of their experiment is immaterial to the point. Ockham’s razor isn’t about demonstrated facts, it’s about unnecessary explanatory principles.Â
By whatever means, they demonstrated to their own satisfaction that the monkeys learned that A comes before B, and that the monkeys learned that B comes before C. Then by like means they verified that the monkeys also knew that A comes before C. This is called transitive inference, and is considered a form of reasoning. Let’s just assume that these are empirically valid observations. My quarrel is not with their methods or results. My quarrel is with my hope, when I saw this, that it might be an example of program-level control. My argument is that it is not.
They are saying, in effect, that the monkeys are controlling the temporal sequence A→C (where → symbolizes the before-after relation) by means of controlling the following series of perceptions at the program level:
For any perceptions A, B, and C
if A→B & B→C
then A→C
I have a real problem with the brain maintaining abstract variables that can be instantiated by any arbitrary perceptions. Bill got into this in part because he thought that’s how language works. But set that aside for another time.
‘A before B’ and ‘B before C’ are sequence perceptions. (You could call them temporal relationship perceptions, except that we privilege time differently from the other three dimensions.)Â
The monkey perceives A→B. in doing so, the monkey necessarily perceives B. But B is the input for perceiving B→C. Here, my experience examining subjective perceptions has suggested to me a more general role for imagination than we have usually considered. I believe that when the monkey perceives B, the monkey has at least a weak imagined signal from comparators that are linked to B in a sequence structure. The sequence A→B is learned and established. Perceiving A, the monkey imagines B. There is an ‘expectation’ of B. There is a readiness to perceive the presence of B when perceptual input is somewhat less complete than ordinarily might be required, because that input is supplemented by an imagined signal. B→C is learned and established. Perceiving B, the monkey imagines C. There is an ‘expectation’ of C. There is a readiness to perceive the presence of C. Consequently, linking up the sequence A→B→C would be a simple and obvious bit of learning. Then to control the perception that A comes before C …
… all they have to do is remember the sequence A-B-C and control it in imagination.
I don’t deny that there are if-then test functions in the neural hierarchy. That’s what input functions do, at every level. All that is needed for the if/then of the above syllogism is the input functions that test for the presence of A, B, and C, the organization of the comparators for recognizing and controlling them into sequences, and a little bit of imagination.
BN: The fact that some behavior can be expressed in formal logical terms doesn’t necessarily mean that’s what the brain is doing. Similarly, just because it can be coded as a program doesn’t mean that the brain is controlling by means of programs. Looking at the oceanic periphery of the PCT map, there be grand delusions here.
RM: PCT doesn’t say that the brain “controls by means of programs”.Â
On pp. 161-169 B:CP proposes that it does. Bill describes “reading the newspaper” as a program that is interrupted by another program, “looking for my glasses”. The latter is “a list of relationships brought about in sequential order”–a sequence. “Yet the program is not a list. It is a structure, and at the nodes of this structure are tests or decision points. There was no way I could have predicted the list of relationships or at what point the list would terminate and a new sort of relationship-list would begin to unfold (when the glasses turned up).” I have a lot of questions and problems with Bill’s description of programs. “Picking up a stiff piece of cardboard requires a different set of lower-order acts than picking up a limp undershirt that happens to have one end caught in a drawer.” This reception of perceptual input by some input functions better than by others, and this varying of perceptual inputs (by varying the reference signals for controlling them)–this is a program? But these questions and problems will have to wait for another occasion. Suffice to say that PCT does indeed say that the brain controls by means of programs, at least PCT as outlined by Bill in B:CP on pp. 161-169 says that.
We have sequence perceptions. Like any perception, a sequence perception can be controlled. A sequence perception comprises a number of controlled variables which are controlled one after another in serial order. Generally, what we are controlling by those means is ultimately the final CV of the series. (There are exceptions to this and elaborations of this.) When we control the final CV of a sequence by means of controlling each of the prior CVs in turn, we have controlled that sequence perception. The final CV might reflect the cumulative effects of the successive CVs, that’s another possibility.
We can perceive a sequence that is controlled by a different agent. Whether this agent is a control system or some kind of non-control mechanism is immaterial. In that case, we perceive a sequence, but we’re not controlling it. We could watch a traffic light cycling through green, yellow, and red, then green again. In this case, we’re in observation mode. The Honeywell traffic light apparatus is controlling the traffic lights, but it is not a control system. It’s governed by a timer. We can imagine a situation where the Honeywell apparatus is on the blink, but until a part can be delivered a technician can stand there and manually switch the lights. The technician is controlling a perception of the correct sequence, acting as a control system. They may refer to a watch but they’re not governed by it. Now suppose the defect in the Honeywell apparatus sporadically swaps yellow and green. Yellow for 3 minutes, then green for 20 seconds, then red for 3 minutes. The technician has to watch and intervene if this happens. If yellow follows red, if green is too short–there are several clues that intervention is needed to switch the apparatus back to proper operation. The technician is controlling a perception of the proper sequence, but the means of control could be anything–bang a fist here, touch a jumper wire across these points in the circuit. They might even have to control “a number of variables one after another in serial order,” perceptions quite distinct from the three colors green, yellow, and red, as the trick to restore orderly operation. So ‘controlling the perception of the sequence’ in that sense is quite different from sequence-level control. In sequence-level control we do not intervene into a sequence performed by some other agent, like Marian the Librarian in “The Music Man” when she moves the child’s finger to the right key, in sequence-level control we are controlling “a series of perceptions in serial order”, do re mi fa sol. Get your hand out of the way, I want to do it myself.
There is a like distinction between controlling a program, as to whether or not it is producing the proper outputs in the proper sequence, and program-level control.
RM: It [i.e. PCT] says that the brain controls program perceptions. And my program control demo shows that this is indeed the case.The demo shows that control of a program perception can occur without the person producing programmatic output. It also shows that control of a program perception is not the same control of a sequence perception. And it shows that program perceptions are controlled at a higher level than sequence perceptions. This is evidenced by the fact that sequence perceptions can be controlled at a much faster rate than program perceptions using when both are controlled using same means (a bar press).Â
Ockham came knocking on your door too. PCT can account for this without invoking the additional level. Your demo says nothing about control at the program level. The delay appears to me to be due to increased complexity across sensory modalities.Â
Arguendo, let’s look at your claim. The demo user perceives when a program has stopped running and presses the spacebar to start it again. So you say the user is perceiving a program and is controlling that perception of the program. The user creates an input function that receives perceptions of the screen outputs of the program that you wrote and which is running in the computer
From these perceptions of outputs on the screen that input function creates a perception of the program. The user had some prior help from your ‘what to look for’ instructions, where you spell out the program code in ordinary language, but even so developing that input function takes time and is not a simple or quick matter. But when the learning process is successful, that input function constructs a perception of a program.
I can think of many experiences that seem similar to this, where I monitor the output of some process and take a prescribed action when it lapses, or I notice in passing that the outputs aren’t as they should be, or someone calls my attention to it, and I take the action that re-establishes proper operation of that process. For that, I just need to know what the outputs should look like. But you’re positing that in order to recognize a lapse, in this case of your demo, I have to be running a replica of the program in my brain: “if circle, then blue; else red”.Â
Th seems to be the necessary structure of the input function for recognizing when the program is running and when it is not, “if circle, then blue; else red”. The input function tests “is that a circle?” and if the answer is yes the input function tests whether the next input is blue; and if the answer to the “circle” question is no the input function tests whether the next input is red.Â
So far, the input function to the program perceiving system ‘PPS’ looks like two sequences (temporal relationships), one where a circle is followed by the color blue, and the other where some other shape (a square) is followed by the color red:
Circle → Blue
Square → Red (or alternatively [error output from the circle detector] → Red)
So the reference value for the program-perceiving function requires a perceptual input from either of these sequence-control systems to know that the program is running. But it also needs to know when the program has stopped running. That’s what produces an error signal which is transformed to the user pressing the spacebar.Â
It would be parsimonious to send the error output from either of these sequence-control systems to the error output of ‘PPS’. But if we’re going to do that, why bother with the error output from ‘PPS’ at all? Just control those two sequence perceptions concurrently. Error output from either of them suffices.
The Circle → Blue sequence is equivalent to “if circle then blue”. The “if” operation is performed by the input function for the first CV of the sequence, and the “then” operation is in the structure of the sequence, such that having perceived the first CV you control to perceive the second; similarly for the other sequence, Square → Red or [error output from the circle detector] → Red. There is no separate ‘if’ test. That’s what input functions do. The contingencies are provided by the structure of a sequence control function and the contingent input from the environment.Â
Where does the perception of the program reside? In your conception, it’s a perception that the program is running correctly (or not). That’s the technician watching the blinking lights. The perception of a program in that sense is not the same as program-level control. Add some contingent perceptual inputs, like a video feed from an overhead drone showing how far traffic is backed up in each direction, and a technician manually switching the lights is probably controlling at the program level.
But in your demo is the task to perceive that the program is running, or is the task to perceive when the program has stopped running? That simple change of sign makes it a game of whack-a-mole.Â
Circle → Red → Whack!
non-Circle → Blue → Whack!
Like the technician waiting for a part at the busy intersection, I want to know when to bump the system back into line. The instruction to in effect run the program in my head in parallel to its operation in the computer strikes me as being like the old-timer counting sheep as they came out of the corral. "How can you count the so fast? What’s your secret?"Â “Oh, it’s easy. I count the legs and divide by four.”
But that just means I’m not a good subject, I guess.Â
···
/Bruce
On Sat, Aug 10, 2019 at 8:47 PM Richard Marken csgnet@lists.illinois.edu wrote:
[Rick Marken 2019-08-10_17:39:08]
[From Bruce Nevin (20190810.09:48 ET)]
https://advances.sciencemag.org/content/5/7/eaaw2089
Reward associations do not explain transitive inference performance in monkeys
RM: If you get a chance could you give a brief summary of what they had the animals doing. And why you think the animals weren’t doing what the researches thought they were doing.
BN: The fact that some behavior can be expressed in formal logical terms doesn’t necessarily mean that’s what the brain is doing. Similarly, just because it can be coded as a program doesn’t mean that the brain is controlling by means of programs. Looking at the oceanic periphery of the PCT map, there be grand delusions here.
RM: PCT doesn’t say that the brain “controls by means of programs”. It says that the brain controls program perceptions. And my program control demo shows that this is indeed the case.The demo shows that control of a program perception can occur without the person producing programmatic output. It also shows that control of a program perception is not the same control of a sequence perception. And it shows that program perceptions are controlled at a higher level than sequence perceptions. This is evidenced by the fact that sequence perceptions can be controlled at a much faster rate than program perceptions using when both are controlled using same means (a bar press).Â
Â
BN: My view:
Logic is a formalization of inferences which in practice we make by other means.
Formal logic can reach correct conclusions which are counter-intuitive.
If logic were innate, logicians would be out of business.
Logic is dependent upon and derived from language.
Frege’s Laws of thought would better be titled Laws for thought.
RM: I think this is pretty much consistent with PCT. In PCT, formal logic is not innate, any more than mathematics is innate. What is innate (presumably) is the ability to perceive the perceptual variables that make up formal logic (such as inferential relationships between perceptions).Â
BestÂ
Rick
Â
Of course this is well-tilled ground where sensible people and I suppose angels should fear to tread.
It’s often overlooked that the truth of a logical conclusion is utterly dependent upon every one of the premises and assumptions of the logical argument. If someone like Alan Dershowitz or Noam Chomsky is beating you up with his impeccable logic, look closely into his premises and assumptions.
/Bruce
–
Richard S. MarkenÂ
"Perfection is achieved not when you have nothing more to add, but when you
have nothing left to take away.�
                --Antoine de Saint-Exupery