While working on a paper on neural computation, the following question came up: can you exert feedback control without processing information?
The kind of case I have in mind is that of relatively simple feedback control devices that use one or two physical variable(s) to affect another. For instance, Watt governors or float regulators that maintain the level of a liquid within a container at a stable level by opening or closing a valve based on a flotation device linked to the valve. When the valve is open, liquid comes in, raising the flotation device and shutting the valve. When the valve is shut, liquid stops coming in. If the liquid goes down again (e.g., because someone flushes the toilet), the flotation device goes down, the valve opens up, and liquid starts pouring in again.
It seems to me that in these control devices there is nothing that deserve the name of information processing. That is, no variables are processed based solely on the information they carry as opposed to their specific physical characteristics. If you want to fix a malfunctioning float regulator, you can’t intervene solely on a hypothetical information-processing part of the process; you just have to adjust the variables based on their physical interaction (e.g., you might change the angle between the arm that holds the flotation device and the arm that opens the valve). Or to put it yet another way, there is no transduction of physical variables into medium-independent variables.
If this is right, then feedback control does not entail information processing (in the general case).
At the same time, if you need to exert feedback control functions of serious complexity, involving the interaction of several input and output variables, the only efficient way of doing so seems to be to transduce all the input physical variables into some internal variables that can be manipulated based on the information they carry and process those. Thus, feedback control beyond a certain complexity requires information processing.
I’d be very interested in any comments that anyone might have. What am I missing?
Hi Gualtiero,
It seems to me that the question depends on how we define information, and how we conceive of an information processing system. Here’s what I wrote in the NATURE NETWORK FORUM in the thread: *What is Consciousness?*
“This is my take on the concept of information.
My definition: Information is any property of any object, event, or situation that can be detected, classified, measured, or described in any way.
1. The existence of information implies the existence of a complex physical system consisting of (a) a source with some kind of structured content (S), (b) a mechanism that systematically encodes the structure of S, (c) a channel that selectively directs the encoding of S, (d) a mechanism that selectively receives and decodes the encoding of S.
2. A distinction should be drawn between latent information and what might be called kinetic information. All structured physical objects contain latent information. This is as true for undetected distant galaxies as it is for the magnetic pattern on a hard disc or the ink marks on the page of a book. Without an effective encoder, channel, and decoder, latent information never becomes kinetic information. Kinetic information is important because it enables systematic responses with respect to the source (S) or to what S signifies.”
So in the case of feedback control of water level by a floatation device, (a) the level of water in the tank would be the latent source of information S, (b) the height of the floatation bulb would be the encoding of S, (c) the floatation lever would be the channel that directs the encoding of S, and (d) the water valve assembly would be the mechanism that receives and decodes the angle of the floatation lever.
In this conception of information and information processing, feedback control is an example of kinetic information processing.
Would you agree? If you disagree, I’d be very much interested in your reasons.
Interesting idea. Indeed, on first thought, this seems to be the right sort of boundary between information processing systems and physical systems…
But here’s an interesting test (it is not definitively, but worth considering). For something to be an information processor, it must perform certain processes over representations. And things that can represent can also misrepresent (perhaps God excluded). So one way to see if something potentially can process information is to see if it can misrepresent.
Can a toilet plumb misrepresent? It would seem so– suppose the plumb/lever got stuck so that it never registered that the tank was full, and as a consequence the toilet over-filled. Sounds like a case of misrepresentation, albeit it a simple one (notice how natural it sounds to say “register”). But the physical story works just as well, and perhaps that is all the plumber needs to do her/his job.
I guess this sounds a bit like Dennett’s intentional stance (maybe there is an “informational processing stance”?). The idea would be that, potentially, any system with feedback could be an information processor. But 1) there are some systems that can also be described physically, and this is a more useful way of predicting, explaining, and controlling their behavior, and 2) there are other systems for which the informational stance is invaluable for P, E, and C.
Bearing on your original question: feedback control, on this kind of view, would not entail information processing or vice versa. But perhaps systems with complicated sorts of feedback control loops are precisely the kinds of systems where explanation from the informational stance are most useful/ indispensable.
thought?
Arnold, thanks for your insightulf comment; I think I agree with what you say. It still seems to me that there is no _processing_ of information in this case, though. There is transmission of information, but no true processing. Or in any case there is no processing of information as such (as opposed to the processing of some particular physical embodiment of the information). What do you think?
Brandon, thanks for your interesting comment. A few points:
1. I don’t like Dennett’s intentional stance framework because it’s too instrumentalistic for my taste.
2. Perhaps the plumb/lever misrepresents and thus carries information; but even if that is a necessary condition for being an information processor, it is not sufficient (nor did you suggest that it is). So the system in question does carry information and perhaps represents, but still does not process information (qua information).
3. I agree that in more complex cases of feedback control, surely information processing is involved.
I don’t know why you want to deny that representing is sufficient for information processing. If some state of the system to be represented has some information, which is represented by the state of some other system, then the information has been processed insofar as its representational matter or form has changed. (Incidentally, with respect to your reply to the first comment on this thread, I think direct transmission *is* a form of information processing.) In any event, the governor and the water-level indicator are even more obviously information processors in that they not only represent information about or in the target system, they do something with that information. Hence, the information in the target system is transformed twice: once into a representation (the angle of a lever-arm) and once into a suitable response action (the opening/closing of a valve). Or if you don’t like actions, take the state of the valve itself as a representation of the target system.
Maybe I’m just not understanding what you have in mind with “processing” (especially with respect to variables), “transduction”, or “medium-independent variables”?
Gualtiero, I’m not sure I understand what you mean when you say “there is no processing of information as such (as opposed to the processing of some particular physical embodiment of the information).” Are you asserting that the “information as such” is separate from its physical embodiment? This strikes me as a crucial point. Can you clarify?
I certainly don’t mean that the information is separate from its physical embodiment. But it seems to that there are two relevant ways of “manipulating” an information carrying variable: (1) we just let the variable affect some other variable (as in our float regulator); (2) we feed the variable (perhaps after transducing it into an appropriate medium) to a processing device whose function is to manipulate the tranduced variable so as to do something useful with that information (such as combine it with other sources of information, or extract some aspect of the informationm, as in a digital computer). Do you think there is no useful distinction here?
Jonathan, thanks for your comment. Your reaction (as well as the reaction of Brandon, Arnold, and others who have emailed me) convinces me that many people are happy to use “information processing” more broadly than I do. So be it. Nothing substantive hinges on it.
That being said, I would still like to distinguish between information processing in the broad sense (in which even transmitting information or encoding a variable is a kind of information processing) with information processing in the stronger sense in which computing over information-carrying signals is information processing. The latter sense of “information processing” is stronger because the information is first transduced into a common medium whose function is to be manipulated by the computing system and then processed by procedures designed with the sole function of extracting or combining the information in useful ways. Would you agree that there is a meaningful distinction there?
I’m not sure that I see the distinction you want. Is the distinction about (1) the medium of representation, (2) whether or not information is manipulated, or (3) the multiplicity of the inputs?
The following operations might correspond to the three points above: (1) count how many and store a representation of that number; (2) count how many, multiply that number by two, and store a representation of the result of the multiplication; (3) count how many in group one, count how many in group two, add those numbers and store a representation of the result of the addition.
It looks to me like the simple feedback systems are carrying out simple computations. If your claim is that there is something special about the medium in which those computations are carried out, then I think I disagree. If your claim is that these simple computations are too simple to count as information processing, then I suppose there is a useful distinction to be made, but it will not have anything to do with the medium but with the complexity of the operations — a purely logical problem.
Hi Gualtiero,
Insofar as we want to avoid casting van Gelder as the heir to Newell and Simon, I think there is a meaningful distinction to be made here. Do you think you’d get any ‘no’ answers if the question were “can you exert feedback control without manipulating items of a representational code/scheme?” My sense is that everyone would say ‘yes’.
I wouldn’t. 😉 If the question were “Can you have feedback without representation?” then I’d say, “Sure.” But insofar as you are talking about *control*, the answer to your question should be “No.”
Jonathan, my distinction has to do with whether an information-carrying variable is transformed simply (1) by letting it have an effect on some other physical variable or (2) by manipulating it so as to do something useful with the information it carries (and regardless of its particular physical implementation). Here’s another way to put the point: to fix or improve the performance of a Watt governor a float regulator you cannot simply improve the algorithm/program/procedure for processing the information while leaving the physics unchanged. But when you put a computer (propery so called) in charge of regulating something, you can. Here’s a third way to put the point: it was a major advance in control theory when people realized that you could use computers to do the information processing that controlled the device, instead of having to build a physical system with the variables related in just the right physical way.
Martin, thanks for being so sensible, as usual! I am no big fan of van Gelder’s work, but I think we should grant him at least that _some_ feedback control occurs without information processing/computation. We actually gain something in the strength of our view that the nervous system processes information.
(If every physical system processes information, than the brain does too. But it’s a trivial conclusion.)
Very good! So, let’s consider the Watt centrifugal governor for controlling engine speed. van Gelder denies that any representations are manipulated, yet he wants to say that it’s a feedback control system. If you are correct, then either the Watt governor does manipulate representations (elements of a representational scheme/code) or it isn’t an instance of feedback control system (or both). Which do you want to say? Might the problem be with how we are using the phrase ‘feedback control’?
I agree that there are many useful distinctions that can be made among different kinds of information processing mechanisms — differences in the complexity of the mechanisms, differences in their operational principles, differences in their composition, etc. What I’m not sure about is where one would draw the line between “weak” information processing and what you call “strong” information processing where there is _computation_ over information signals.
Your example of strong information processing is one where “… we feed the variable (perhaps after transducing it into an appropriate medium) to a processing device whose function is to manipulate the tranduced variable so as to do something useful with that information.” It seems to me that the transduction of latent information into a form that can be used by the mechanism corresponds to an encoding of information, and the decoding of the encoded information is in fact doing something useful with that information, i.e., the systematic transformation of latent information into kinetic information. This is what happens in the feedback control of water-level in a tank.
Is this, in principle, different from what happens (e.g) when we compute 4+9 in a digital computer? In this case, pressing “4” then “+” then “9” on the keyboard creates the source of latent information; this latent information is then encoded into logic-gate settings within the computer; the internal binary machinery of the computer operates sequentially on these settings to perform an add function and decodes the binary output to yield the number “13”, which is kinetic information that enables an appropriate response by the typical human user. If, instead of a modern digital computer, all of these operations were performed by an old-fashioned analog computer (relays, motors, gears), the gap between the automatic addition of numbers in the computer and the automatic closing of a valve by feedback in the water-tank example would not seem so great. Difference in complexity? Yes. Difference in basic principle? It doesn’t seem so to me. Of course this doesn’t preclude your point that useful distinctions can be made among different information processing systems.
I apologize for being so slow to understand, but I’m still not getting it. I have some dim sense of what it means to let a variable affect some other physical variable. For example, if I have a variable M for mass and a variable L for length, I can think about M causing L according to Hooke’s law (an equation like L = L0 + kM + error).
I’m not at all clear about what you mean when you say that a variable is transformed by “manipulating it so as to do something useful with the information it carries (and regardless of its particular physical implementation).” Specifically, I don’t understand two things that you say. First, what do you mean when you say that some variable is manipulated? Do you have in mind Pearl’s DO calculus or something else? If something else, what? Second, what do you mean when you say that the use of the transformed variable (or is it the manipulation itself?) has no regard to its particular physical implementation? You seem to agree that information must have *some* physical basis. So, any instantiation of information–in a Watt governor or in a computer–must have some physical properties that matter. Moreover, in both computers and governors, there will be some physical properties that don’t matter. Governors work regardless of their color, for example. In fact, lots of different designs for governors have been implemented, and some are more efficient than others (sort of like some programs are more efficient at using physical resources, like memory and clock cycles, than others).
By your third point, I’m no longer sure I understand what a physical system is. I would have said that a computer is a physical system and that a computer would have to be related in just the right way to any system it was going to represent. Is that an unfair/unhelpful thing to say? If so, why?
I’m not sure what it means to manipulate a representation. What I am reasonably confident about is that some parts of the governor represent engine speed and then use that representation in order to regulate engine speed. I want to say that the engine speed is encoded in the speed of the flywheel and transmitted to the angle of the flyballs. The angle of the flyballs then effectively sends a signal to a valve telling it when to open and when to close and by how much. (If I recall correctly, that signal is sent via a clever arrangement of rods of some sort.) I think this sits very comfortably in the language of representation scheme and/or code.
What I meant to indicate by saying that there is a difference between feedback and feedback *control* is essentially this: with feedback, it might be that no one notices or cares about the feedback, and so, even though some variables in the system contain information, they do not function as signs for anyone. In control systems, someone has set out with a problem. He or she explicitly wants to *control* some aspect of a system. Hence, that person uses some variables corresponding to some physical features of the system to represent other physical features of the system. I just don’t see how anything could be said to be *controlling* some feature of a system without representing that feature.