The problem with the theory of representation Chemero settles on in the previous chapter, as we shall see, is that it makes representation unavoidable. If there are representations, then RECS fails to get off the ground (remember, the radical bit is anti-representationalism). In this chapter, Chemero identifies a path past this problem (the 'dynamical stance'), explains why it works, but then concludes with a new problem, that of how to guide future discovery.
Chemero identifies two levels at which one can deny representations. The first is metaphysical, that is, you can deny that there is anything like a representation in any cognitive system; this is a philosophical hypothesis. The second is epistemological, where you simply claim that none of our theories of cognition need feature representations, without taking a stand on what cognitive systems are actually made from. This is more a scientific hypothesis. Chemero will argue that the former argument is indefensible, because of how wide-reaching the definition of representation is, and will advocate defending the latter epistemological hypothesis, suggesting radical embodied cognitive scientists should simply adopt a dynamical stance in our theorising and experimentation. Philosophers of science should recognise the origin of this approach as Daniel Dennett's intentional stance, where you simply work from the assumption that an agent has beliefs and desires and, on that basis alone, quickly and efficiently predict it's behaviour without worrying about the details of the mechanism (Dennett, 1987). In the same way, Chemero is advocating that we should be satisfied with the kind of complete, counter-factual supporting dynamical models that are becoming more common in cognitive science, and not try to add a representational gloss.
Watts governors & evolutionary robotics
van Gelder (1995) described the Watts steam governor as a classic example of a non-computational, non-representational dynamical system that could serve as a model for the radical anti-representationalist. and support for the metaphysical claim. Sabrina has discussed this paper at length, but to summarise, the Watts governor is a calibrated device for controlling the engine speed of a steam driven engine. According to van Gelder, it is best described as two non-linear dynamical systems (the governor and the throttle valve) which are coupled together to produce the required stable speed. The equation of motion for the governor describes how the angle, θ, between two arms changes as a function of, in part, the speed of the engine ω; the equation for ω in turn,depends in part on θ. The two equations are thus coupled (non-linearly, as it happens).
There are two representational accounts of this system; one is computational (van Gelder describes a computational algorithm for changing the speed if it is found to be incorrect but rejects it as not actually describing anything present in the system), and one simply claims that the arm angle θ represents the speed of the engine. van Gelder resists the latter account entirely, but Chemero is convinced that, according to the theory of representation from Chapter 3, θ is a representation of ω and thus there is a legitimate representational account of the governor. Given the existence of this account, the Watts governor is not a slam dunk case for the metaphysical claim that there are no representations.
Chemero is happy, however, that it is support for the epistemological claim, because it is not clear the representational story adds anything to the dynamical account. Critically, the dynamical account must come first; you can't tell a traditional representational story without some idea of the function of the system, which in this case comes from the dynamical account. Given that it doesn't add anything, you might simply wish to stop with the dynamical account and not concern yourself with the representation that is in the system; you could, in Chemero's phrase, simply take the dynamical stance towards the system in question. (This analysis is the topic of a paper by Nielson, 2010.)
Like the intentional stance, the dynamical stance is only worth taking if it works. Chemero therefore notes that the dynamical stance is worthwhile if and only if there a) is a large class of dynamical models which don't need a representational 'gloss', and b) these models include the best explanations for cognitive phenomena. The latter is an empirical question; Chemero now discusses an example that supports the former requirement, evolutionary robotics ('Sussex robots').
These robots are allowed to evolve, with the only selection pressure success at their current task. The strategy is to allow the robots to find a solution to the problem without having the solution programmed into them. The Sussex group describe the robot and environment as coupled dynamical systems. For example, a robot evolved to track targets can be fully described as a simple recurrent feedback network. It's environment is then fully described as a state space defined in terms of egocentric polar coordinates - all you need to know is the distance from the robot to the centre of the space, and the clockwise angle it needs to turn to face it. (Creating a space like this is directly analogous to what we do to characterise coordination, where we use relative phase as the 'world' variable.) These two dynamical systems are then simply coupled, and the required dynamics (tracking behaviour) then emerges from this coupled system in the form of a point attractor on the target (see Harvey et al, 1994).
Under the traditional theory of representation Chemero is targeting, there is still a representational account for this robot. The system contains visual input nodes ('representation producers') which produce activations across intermediate nodes ('representations') which affect the behaviour of motors via three other nodes ('representation consumers') to produce the tracking behaviour ('adapting the system to some part of the environment'). But Chemero describes (p.77) how this representational gloss doesn't help - it could only be constructed after we had the dynamical account, and the dynamical account already provides a complete characterisation of all possible behaviours: we can use it to predict behaviour with no reference to the representational story. Taking the dynamical stance has 'paid off', and while it remains an ongoing task for dynamical systems cognitive science to actually produce these types of models, there are already numerous examples in the literature of dynamical accounts of complex behaviour which make no reference to representation.
Guides to discovery
So the dynamical stance seems to be a sensible epistemological approach - we can fully describe and predict behaviour using dynamical systems terms without any mention of representations, and while there may indeed be a representational story for every dynamical one, they do not necessarily add any descriptive or explanatory power. So we can do cognitive science without representations.
Or can we? For better or for worse, the theory that cognition is representational (and, in general, computational) has been a productive theory because it is, actually, a theory in the scientific sense. Scientists can use it to make predictions, and design experiments to test the prediction. In other words, representational theories of cognition serve as a guide to discovery. Chemero highlights the fact that the dynamical systems approach is effectively phenomena driven - you find a behaviour, and characterise the system with a model, but it's all after the empirical fact: the model didn't suggest the next experiment to run. This is a real problem - historically, it mirrors a fight in physics between the theory of unobservable atoms and Mach's phenomenology, which Mach lost entirely and for the good reason that atom theory could drive experimental work while phenomenology cannot. Regardless of whether there actually are representations, it would seem that the most productive way to progress is to act as if there are, so that the field can progress.* If this is the case, then even the (comparatively weak) dynamical stance is the wrong move, and RECS, again, fails to get off the ground.
The next section of the book lays out two suggestions: Chemero will propose that a) actually, dynamical models can, in and of themselves, serve as a guide to discovery, and b) a version of Gibson's ecological psychology will also provide the theoretical tools a radical embodied cognitive psychologist will require.
Some thoughts
It's about this point that I really start to worry about the definition of representation that Chemero is using; it is apparently so broad as to start losing it's explanatory value. This may be the case, although Chemero did go to great lengths in the previous chapter to avoid setting up a straw man. But if this is what representations are, I'm happy to not be using them.
I am not actually convinced that the representational accounts of, say, the Watts governor make any sense at all. These accounts focus on the arm angle as a representation of speed; Neilsen (2010), for instance, tries to make the dynamical account representational by rearranging the equation of motion so that it is expressed with the arm angle on the left hand side of the equation. This is certainly possible, but a) hardly convincing given b) the original differential equation is presumably written with respect to the spindle for a reason, namely that is what is to be controlled. I think I take Chemero's point about this being a mere representational gloss fairly seriously: it's just not helping, at all. That said, I think 'stances' (intentional or dynamical) are fairly weak as well, so I'm more interested in arguments that enable me to avoid representational talk for reasons other than convenience.
The next sections are designed to provide some tools along these lines. The dynamical model Chemero is going to use as an example of a guide to discovery is the HKB model. This is an odd move, given that a) Kelso explicitly described his approach as phenomenological and b) the discoveries the HKB model guided us to mostly suggest this approach is incorrect. In my next post I'll critique this in more detail.
I am, of course,on board with Chemero's other suggestion, that a) dynamical systems is not a theory of behaviour but that b) Gibsonian ecological psychology can and should be the theory of choice. I'll have plenty of words to say about the specific flavour of ecological psychology Chemero will endorse, though (I've already talked about the issue of affordances).
_______________________
*Of course, one problem with modern psychology is that people have forgotten this distinction.
References
Dennett, D. (1987). The Intentional Stance. Cambridge Mass: MIT Press.
Harvey, I., P. Husbands, and D. Cliff (1994). Seeing the light: Artificial evolution, real vision. In From Animals to Animats 3, ed. D. Cliff, P. Husbands, J. A. Meyer, and S. W. Wilson. Cambridge, Mass.: MIT Press.
Neilsen, K. S. (2010). Representation and dynamics. Philosophical Psychology, 23(6), 759-773. DOI
Van Gelder, T. (1995). What might cognition be, if not computation. The Journal of Philosophy, 92(7), 345-381. Download
There are two representational accounts of this system; one is computational (van Gelder describes a computational algorithm for changing the speed if it is found to be incorrect but rejects it as not actually describing anything present in the system), and one simply claims that the arm angle θ represents the speed of the engine. van Gelder resists the latter account entirely, but Chemero is convinced that, according to the theory of representation from Chapter 3, θ is a representation of ω and thus there is a legitimate representational account of the governor. Given the existence of this account, the Watts governor is not a slam dunk case for the metaphysical claim that there are no representations.
Chemero is happy, however, that it is support for the epistemological claim, because it is not clear the representational story adds anything to the dynamical account. Critically, the dynamical account must come first; you can't tell a traditional representational story without some idea of the function of the system, which in this case comes from the dynamical account. Given that it doesn't add anything, you might simply wish to stop with the dynamical account and not concern yourself with the representation that is in the system; you could, in Chemero's phrase, simply take the dynamical stance towards the system in question. (This analysis is the topic of a paper by Nielson, 2010.)
Like the intentional stance, the dynamical stance is only worth taking if it works. Chemero therefore notes that the dynamical stance is worthwhile if and only if there a) is a large class of dynamical models which don't need a representational 'gloss', and b) these models include the best explanations for cognitive phenomena. The latter is an empirical question; Chemero now discusses an example that supports the former requirement, evolutionary robotics ('Sussex robots').
These robots are allowed to evolve, with the only selection pressure success at their current task. The strategy is to allow the robots to find a solution to the problem without having the solution programmed into them. The Sussex group describe the robot and environment as coupled dynamical systems. For example, a robot evolved to track targets can be fully described as a simple recurrent feedback network. It's environment is then fully described as a state space defined in terms of egocentric polar coordinates - all you need to know is the distance from the robot to the centre of the space, and the clockwise angle it needs to turn to face it. (Creating a space like this is directly analogous to what we do to characterise coordination, where we use relative phase as the 'world' variable.) These two dynamical systems are then simply coupled, and the required dynamics (tracking behaviour) then emerges from this coupled system in the form of a point attractor on the target (see Harvey et al, 1994).
Under the traditional theory of representation Chemero is targeting, there is still a representational account for this robot. The system contains visual input nodes ('representation producers') which produce activations across intermediate nodes ('representations') which affect the behaviour of motors via three other nodes ('representation consumers') to produce the tracking behaviour ('adapting the system to some part of the environment'). But Chemero describes (p.77) how this representational gloss doesn't help - it could only be constructed after we had the dynamical account, and the dynamical account already provides a complete characterisation of all possible behaviours: we can use it to predict behaviour with no reference to the representational story. Taking the dynamical stance has 'paid off', and while it remains an ongoing task for dynamical systems cognitive science to actually produce these types of models, there are already numerous examples in the literature of dynamical accounts of complex behaviour which make no reference to representation.
Guides to discovery
So the dynamical stance seems to be a sensible epistemological approach - we can fully describe and predict behaviour using dynamical systems terms without any mention of representations, and while there may indeed be a representational story for every dynamical one, they do not necessarily add any descriptive or explanatory power. So we can do cognitive science without representations.
Or can we? For better or for worse, the theory that cognition is representational (and, in general, computational) has been a productive theory because it is, actually, a theory in the scientific sense. Scientists can use it to make predictions, and design experiments to test the prediction. In other words, representational theories of cognition serve as a guide to discovery. Chemero highlights the fact that the dynamical systems approach is effectively phenomena driven - you find a behaviour, and characterise the system with a model, but it's all after the empirical fact: the model didn't suggest the next experiment to run. This is a real problem - historically, it mirrors a fight in physics between the theory of unobservable atoms and Mach's phenomenology, which Mach lost entirely and for the good reason that atom theory could drive experimental work while phenomenology cannot. Regardless of whether there actually are representations, it would seem that the most productive way to progress is to act as if there are, so that the field can progress.* If this is the case, then even the (comparatively weak) dynamical stance is the wrong move, and RECS, again, fails to get off the ground.
The next section of the book lays out two suggestions: Chemero will propose that a) actually, dynamical models can, in and of themselves, serve as a guide to discovery, and b) a version of Gibson's ecological psychology will also provide the theoretical tools a radical embodied cognitive psychologist will require.
Some thoughts
It's about this point that I really start to worry about the definition of representation that Chemero is using; it is apparently so broad as to start losing it's explanatory value. This may be the case, although Chemero did go to great lengths in the previous chapter to avoid setting up a straw man. But if this is what representations are, I'm happy to not be using them.
I am not actually convinced that the representational accounts of, say, the Watts governor make any sense at all. These accounts focus on the arm angle as a representation of speed; Neilsen (2010), for instance, tries to make the dynamical account representational by rearranging the equation of motion so that it is expressed with the arm angle on the left hand side of the equation. This is certainly possible, but a) hardly convincing given b) the original differential equation is presumably written with respect to the spindle for a reason, namely that is what is to be controlled. I think I take Chemero's point about this being a mere representational gloss fairly seriously: it's just not helping, at all. That said, I think 'stances' (intentional or dynamical) are fairly weak as well, so I'm more interested in arguments that enable me to avoid representational talk for reasons other than convenience.
The next sections are designed to provide some tools along these lines. The dynamical model Chemero is going to use as an example of a guide to discovery is the HKB model. This is an odd move, given that a) Kelso explicitly described his approach as phenomenological and b) the discoveries the HKB model guided us to mostly suggest this approach is incorrect. In my next post I'll critique this in more detail.
I am, of course,on board with Chemero's other suggestion, that a) dynamical systems is not a theory of behaviour but that b) Gibsonian ecological psychology can and should be the theory of choice. I'll have plenty of words to say about the specific flavour of ecological psychology Chemero will endorse, though (I've already talked about the issue of affordances).
_______________________
*Of course, one problem with modern psychology is that people have forgotten this distinction.
References
Dennett, D. (1987). The Intentional Stance. Cambridge Mass: MIT Press.
Harvey, I., P. Husbands, and D. Cliff (1994). Seeing the light: Artificial evolution, real vision. In From Animals to Animats 3, ed. D. Cliff, P. Husbands, J. A. Meyer, and S. W. Wilson. Cambridge, Mass.: MIT Press.
Neilsen, K. S. (2010). Representation and dynamics. Philosophical Psychology, 23(6), 759-773. DOI
Van Gelder, T. (1995). What might cognition be, if not computation. The Journal of Philosophy, 92(7), 345-381. Download
Posts
This chapter is a little tough for me, but since I'm an undergraduate that's to be expected. What I did find easy to pick up, however, is the distinction between metaphysical and epistemological 'modes of attack' against representationalism. I looked at it as analogous with atheism and agnosticism. In a world where representationalism is God, those who make metaphysical claims are the atheists. Representations do not exist, period. Those who support the epistemological approach are the agnostics. There might well be representations out there, but you can explain things without needing to know for sure if they are real.
ReplyDeleteHmm. The "dynamical stance" is interesting but I wonder if it isn't just hand-waving?
ReplyDeleteFor it seems as though one could apply this kind of analysis to any system and declare that representational accounts are mere gloss, even if the system is, clearly, representational:
take for example a scanner connected to a printer. you put a piece of paper in it, it scans it, and then it prints you a copy 10 seconds later.
the way it does this, is by storing a representation (say in .jpg format).
But you could say, well I don't care about that, I'm taking the dynamical stance and I think that this system behaves as follows: it takes a document and makes a copy.
you could make a lovely set of equations describing exactly how the copy relates to the original, the differences in resolution etc. to allow you to predict the output based on the input.
which is fine and adaquate if you're only interested in how to make a copy of your tax return form. But clearly you have not understood the system. Whereas if you work out that it's a scanner which converts the data into a .jpg and sends it via a USB connection to the printer, you have (partially).
my worry is that the 'dynamical stance' sounds a little like saying, I just don't care about representations. I'm going to proceed without giving them any thought and see how far I get. the problem being that this might take you very far (mapping the scanner's input to its output) while missing the whole point.
Neuroskeptic,
ReplyDeleteI think Chemero deals with this issue pretty well. In fact, he states that one could write a computer program to control the speed of the engine in the way the Watt generator does, AND he states that if the mind does what his computer program would do, then we WOULD want to use representation talk. So, I think he would agree with you that the scanner represents the page when it stores it as a jpg.
I personally think that all these discussions should start with clear and obvious representations, e.g., a painting that re-presents Henry VIII. We would not be allowed to move on, until we all agreed on what we mean when we say that the painting re-presents Henry. Then, with that understanding in hand, we would evaluate the evidence that the brain does something like that. Even though I feel the need to stand up for the obvious meaning of words, I like the way in which Chemero by-steps the issue - it moves the book forward much faster.
@Kelly
ReplyDeleteYes, that's about right. Actually putting in that way gets to the heart of my unease with 'mere stances'; I turned into an atheist rather than an agnostic because I thought the latter wasn't the right kind of stand. Acting 'as if' something is true bugs the scientist in me, even if it's very hard to make the metaphysical case.
@Neuroskeptic
In Tony's defence on this, there's a key section that I mentioned:
Chemero therefore notes that the dynamical stance is worthwhile if and only if there a) is a large class of dynamical models which don't need a representational 'gloss', and b) these models include the best explanations for cognitive phenomena.
So, as Eric notes, the printer case would fail the test because there is a better story lurking, waiting to be found.
Chemero also lays out that, for the dynamical systems in question (Watts governors and Sussex robots) the dynamical story came first and is counter-factual supporting, i.e. if/then kind of logic - they could be wrong and shown to be wrong. That counts as scientific, even if eventually wrong.
But you are right - stances, intentional or dynamical are risky and not, I have to say, super convincing to me as a scientist. I wasn't that into Dennett when I first came across it either.
Once metaphysical claim and epistemological claim were understood the chapter was easier explained. It was something that when real life examples are given they became easier to undetstand and grasp.
ReplyDeleteAlso this chapter was challenging when it came to the math equations, but when we discussed the chapter in the class it was found that we didn't really need to understand the equation, Chemero explained everything through the chapter.
Gaby, that's about the best way to start getting comfortable with the maths; I'm not that much of a mathematician but I can figure out these kinds of equations now because over the years I've looked at them while reading the English explanation :)
ReplyDeleteAndrew and Gabby, the best way by far to understand most mathematical equations are to read the english translation for them. Words problems are by far the most difficult of equations and they don't get any easier as higher level skills of math come to hand. As far as metaphysical and epistemological claim in the chapter go, Kelly's "religion" example we used in class was the easiest way to explain the difference between the two. Chemero's way of explaining was a little disappointing saying basically that the epistemological claim is more or less a scientific hypothesis and "concerns how what the mind is really like" whereas the metaphysical claim is a philosophical hypothesis and needs science to explain. It was quite confusing, though I knew what he was trying to say. The "religion" example was more down to Earth and easier to grasp the concept of.
ReplyDeleteI'm also an undergraduate taking a course where we are reading this book and found this chapter to be challenging. Once a metaphysical claim and epistemological claim was defined I felt I understood the chapter better. In the case of the Watt governor using an example of a theater ball where when the "steam" pushes the ball around until it reaches the opening where and then blocks the "steam".
ReplyDeleteAs an undergrad, I'm finding this entire book difficult to read. Until the class discussions, I am usually lost. During this chapter, however, Kelly's Atheism/Agnosticism analogy really helped me to understand. I mean, I understand the general concept of the Watt governor, but the rest was Greek to me.
ReplyDeleteAre there specific things that aren't clear? I didn't really have an undergraduate audience when I wrote this, for example, so it may not be much help - but are there particular things people are finding hard?
ReplyDelete