Notes on a Lecture on "Origins of an Embodied Cognition: Moving, Perceiving, and Thinking in Infancy"20 May 1996 01:34
Notes on the psychology lecture yesterday (22 February 1996), Prof. Esther Thelen of Indiana University, speaking on "Origins of an Embodied Cognition: Moving, Perceiving, and Thinking in Infancy" (sponsored by the Psychology Department and the University Lectures Committee). The abstract was quite interesting:
Gaining control of the body is the foundation of development. To function int he world, infants must be able to move eyes, heads, and limbs appropriately, stabilize posture, and locomote through space. Since Piaget, it has been widely recognized that these early perceptual motr skills are the foundation of later, more abstract thinking. Yet the classic Cartesian problem remains: how can abstract, representational qualities of mind emerge from origins in bodily taks?After that, it was, frankly, a great disappointment.
A dynamic systems framework restates the problem in terms of continuous processes and nested time scales. The ways in which infants solve problems of moving real physical bodies through space are continuous with the ways they learn to think. The challenges are to match current capabilities with goals through exploration and learning, to remember, and to generalize. There are no discontinuities between these processes in everyday action, and those over longer time scales. Because development begins with problems of body masses and forces, and because we continually think in order to act, cognition must be embodied and dynamic.
In this lecture, I will show how infants' early perceptual-motor problem solving initiates this developmental cascade.
- Hand-waving about dynamical systems and embodied mind --- cites the names of Lakoff, Johnson, Searle, Dennett, Varela [Gag!].
- Doesn't seem to have grasped Dennett, who's very much a computationalist. She did not name Pavlov, Watson or Skinner, but (see on) she could have.
- Claims that a system is something made of parts which show coherent behvaior
- This is surely false: take two systems with no connection or coherence inter se and declare them sub-systems of a new, larger system. (This is a trivial procedure, but it works, and non-trivial examples can be constructed; a nice CA example is the non-linear voter model with all probabilities set to 1/2.)
- Claims that the behavior of the whole is always simpler than that of the parts, and instances the solar system.
- I can't make heads or tails of that. Does she perhaps mean the motion of center of mass? What about an ideal gas?
- Claims that this simple behavior then leads to complexity again
- to which one can only say, ha!
- Phase plot of a spring, claims it was damped
- but in fact the plot showed a limit cycle, so it was not not just damped
- Defines non-linearity as "non-linear response when you change one of the parameters linearly"
- which ain't even wrong, and her intent would, it seems, make parabolic motion a non-linear process! (For the non-physics-geeks: a projectile fired off with some initial velocity, and subject to gravity and no other forces, moves in a parabola, which is not a straight line. But it's motion is the simplest linear dynamical system, a fact which has eluded a number of people.)
- Shows logistic map and bifurcation diagram as an illustration of non-linearity
- Which again isn't even wrong.
- Dynamical systems are continuous [well, continuous ones are! --- CRS] and "non-computational"; no one is computing the weather.
- But then no analog computer is computational either, since it's a continuous dynamical system; and if we include discrete dynamical systems, no finite state automaton or even Turing machine is computational, which is a reductio. Worse, digital computers are hunks of matter under the usual laws of physics, and so, presumably, continuous dynamical systems. (Neither continuity nor non-computationality were used in her presentation, though I think she wanted to use the second, non-computation, to beat those who talk about the brain as a computer with.) --- Anyway, if she's not upset about saying pre-verbal infants engage in problem solving with memory, what's wrong with computation?
- Showed a graph of horse's oxygen consumption vs. speed for three different gaits; these overlap (i.e. there's hysteresis), and horses are most often found near one of the three minima. Claims this follows from a potential function [!] which itself has minima near those points of minimal oxygen consumption.
- (1) No evidence that a proper potential exists; (2) might describe the horse's minima-seeking, but no way no how explains it!
- "Brain is a dynamical system, the body is a dynamical system, the environment is a dynamical system"
- Well, we may grant her this.
- Says that therefore [her therefore! --- CRS] there is no thought or abstract reasoning which is not emboddied, and that anything else is dualism
- This is highly arguable, and may well come down to words. A traditional
materialist might well say that, yes, all thought is embodied, meaning a
thought is something which happens in the brain, and (one way of)
denying that is dualism. Saying that the thought is really some sort
of manipulation of the body and couldn't happen without the body is a separate
question, and that in turn is separate from saying that learning to manipulate
the body in certain ways is a necessary causal antecedent to learning to think
in certain abstract ways. One could, with perfect consistency, say, yes, the
brain is a dynamical system, and yes, our abstract thought must be preceeded by
concrete bodily motion, but "having climbed the ladder, we kick it away"; the
paralyzed can still think, worse luck. Whether one could still think with the
motor cortex damaged is another question; whether this would be
because motor functions are gone, or because some piece of hardware
there has a dual use, is yet another question, all of which she was trampling
over, and using "dualism" as a stick to beat us over the head with --- this, at
least, was my impression, based on her reaction to questions at the end.
Someone asked if she thought the same approach could be extended to language,
where, after all, the raw physics isn't very important, and she said
that only a dualist would deny it. Fah!
More pithily: Origins aren't everything.
- They are initially strongly coupled, in an uncontrolled way
- Perhaps. It remains to be shown. It wasn't, at least not here.
- Adaptive behavior or skilled behavior consists of learning to control the couplings so as to keep the performance of the task from being affected by changes in the environment... (She instances: driving a car in a windstorm; abstract reasoning.)
- This is one part Ashby and one part Simon, in short Miller, Galanter and Pribram, Plans and the Structure of Behavior, 1959. Hell, it's in the first chapters of William James in 1890. But notice this "learning to control the couplings" business (she used the word "tune") --- how much of that is needed, and how much raw control of things like muscular tension? Anyhow, what is doing the controlling, and setting the goals? It seems like she is introducing this goal-setting, learning agent (or agents) which is not in the dynamical system under consideration and computational (the learning part, feedback). And she gets off saying that language theorists are dualists? Feh. But one can see why shes does this, since to do otherwise she would have to somehow include all these parts of the brain's function into the dynamical system, and we know squat about that. In fact, to do this right, you either have to just consider some slice of the body-and-brain, and then have exogenous control imposed (in which case the "embodiment" argument collapses); or explicitly include everything in the system, and show that with the brain included, given the right initial conditions, the brain comes to act in the right way; as part of this you'd have to define things like "adaptive behavior" and "goals" in dynamical terms, which is probably not impossible but is certainly not easy.
- But, because of the original strong couplings between brain, body and environment, there is no such thing as non-embodied thinking.
- Again, this is a total non sequitur. She admits that learning to, say, think abstractly, involves being able to isolate part of the brain from outside disturbances; and what more do you need? Nobody's pumping for a ghost in the machine, really. [On re-reading this, several months later, having read part of Johnson's The Body in the Mind in the interval, I find my own comments less convincing; but I continue to find Thelen far less than convincing. CRS] In any case, a dynamical systems analogy may be helpful: entrainment of oscillators can be reversed, i.e. oscillators can go out of sync, and you can even construct systems where they drive each other out of sync, endogenously.
- Phase transitions are transitions from one attractor to another, and occur when one attractor becomes unstable
- Well, --- sometimes, sometimes, and sometimes. Not all transitions from one attractor to another are phase transitions, since you can have two, co-existing, stable attractors, and a fractal boundary between their basins of attraction, so that which attractor an initial point near the boundary will end up on depends sensitively on the initial conditions, in fact to infinite precision. And there seems to be a great deal of confusion between stable orbits and stable attractors, since they're stable or unstable with respect to different classes of disturbances. And it's hard to square this what happens during, say, a pitchfork bifurcation, when one of the attractors didn't exist before, and what used to be an attractor isn't any more. And there are phase-transitions in non-dynamical systems, like the Ising model.
- At times of transition and instability we are most likely to see the innate dynamics of the system
- This is nonsensical, given her account. She is saying that the brain, the body, and the environment are all dynamical systems, and are all coupled together; therefore they are parts of one dynamical system, and we are always seeing the intrinsic dynamics. Whether we choose to say "oh, those state variables are associated with the brain" or not, is merely a matter of notational and perhaps cognitive convenience; it makes no difference to the system at all.
- Development is guided by some kind of potential function over the space of possible behaviors [and bodies? --- CRS], and over time some behaviors (like crawling) first become attractors, and then repellers
- In a sense, this is not novel at all; the "developmental landscape" or mountain goes back, in embryology, to Waddington in the nineteen thirties, if not before, i.e., this idea is sixty years old at least. But Waddington was merely being metaphorical, and knew enough not to postulate an explicit potential function. The mathematical difficulties involved in that are formidable, and don't seem to be appreciated. (You can't find a potential function for most dynamical systems.) --- In any case: what controls the evolution of the potential function?
- So, when we look at the transition from one type of behavior to another, we can see the intrinsic dynamics of the system best
- Again, the "intrinsic dynamics" line is inconsistent with her overall position, and in any case unclear. When the infant is learning how to do something new, we expect failures, confusion, reversions to the old behavior, etc. It is not at all clear that the new version provides an explanation of these, rather than a new description, particularly since the models proposed are not detailed ones, but rather extracted from experimental data, and fitted to them.
- The first examples have to do with the uncoordinated motion of infants, before they learn to crawl, grasp, etc., and are just kicking. These, it is claimed, are rhythmic, with spring-like properties.
- Phase plots (x vs. dx/dt) were shown. The paths overlapped themselves, so either (a) system parameters were changing as time went on (b) there were other variables or (c) both. Claimed these were driven oscillators. Perhaps, but at the lecture you'd have to take her word for it. Also claimed that the infant was learning the parameters of the muscles. Again, perhaps.
- Claims to illustrate adaptation and skill as tuning the parameters by showing infants walking on treadmills, including a pair of treadmills, one to a foot, moving at different speeds
- All this shows is that an infant has to be able to control its legs to walk, and that it can learn to move the legs separately. This is merely descriptive, and tells us nothing about how it is done. At some level, parts of the infant's nervous system must be adjusted to the mechanical characteristics of its limbs, but saying this is a dynamical system doesn't help explain how this is done at all. So really this doesn't tell us very much at all.
- Next turns to grasping, and infants who are just learning how to grasp objects. Observes that they make unsuccessful attempts, and claims that these are the intrisic oscillations of the system, unmodulated.
- I say the poor beasts are trying are trying to grab the objects, fail, and try again, until they either succeed or give up. Now these "motion stereotypies" may be the result of some kind of oscillation in the nervous system, but that certainly wasn't shown by the video tapes, and (again, again, again) if you are not a dualist, what makes these motions intrinsic, but the goal-directed, successful attempts to grasp extrinsic?
- Piaget's A not B task. Take an object, put it in a hole, cover the hole with the lid. Infants can get it from under the lid. If you use the same hole a number of times (A), and then, in their sight, hide it under lid B, the infants will (almost always) open hole A, and be surprised. Piaget interpreted this as showing that the infants, at this stage, do not have an idea of the persistence of objects in space and time. She claims that what's really happening is you're training them to reach for A, and that you don't need to hide the object; that they are more likely to reach for A, the more often you have them do it; that moving them or distracting them visually reduces the odds of their going for A to pure chance, or nearly that. Illustrates this by a potential well at A, which gets deeper and deeper. Claims that the task leads to the problems because it is novel and confusing. The motions the infant makes are jerky, and after a number of training trials the jerks freeze into a rather repeatable pattern, i.e. the v vs. t graphs are converging to a certain function.
- I'm not about to dispute her experimental results, but look at what she's saying: you build up a habit, in the form of associating a certain sequence of muscular movements with a certain class of stimuli. The infant is then presented with a similar stimulus, and it engages in the associated habit. Psychology, I believe you will remember Profs. Pavlov and Watson... The dynamical systems business and especially the potential function adds absolutely nothing to our understanding of this behavior, though it's interesting to know that Piaget was wrong. It's also interesting to know that the infant converges on a certain series of muscular movements, though whether this tells us something about the feedback properties of the system, or just that the infant learns to associate the chain of motions a, b, c, etc. (à la William James, one of the earlier chapters again), is unclear.
- She concluded with some more handwaving about development, presenting two overheads, one of which was supposed to be a Piagetian constructivist view, another a [Chomskyan] "amazing baby" view, where pre-programmed information just unfolds at certain stages, and dismissed them both as "dualist" in favor of an "embodied" view, of which no details were given.
- But this is absurd. At some point we do begin to think abstractly, even if we don't do it very well; and evidently quite detailed developmental programs can unfold in other parts of the body, such as the immune system, or the heart, or the lungs. So either development is not purely a dynamical system, or at dynamical systems can, using genetically encoded information (i.e., from certain initial conditions) unfold very detailed structures, which could very well include modules for language, vision, naive physics, etc. In fact, it could even include modules for abstract logic, since that has some survival value. Now these may well need, say, the motor-control parts of the brain to be already in place (as the formation of the heart needs differentiation into muscle tissue), and may well need certain sorts of interaction with the environment to turn out well (but we're including the environment in the dynamical system, so this is purely endogenous); but that's all. (I am not familiar enough with Piaget's work to say whether it could be defended along similar lines.)
Thelen has published a book, A Dynamic Systems Approach to the Development of Cognition and Action (co-authored with Linda B. Smith) which I ought to read, since she may just have been having an off day, or have pitched things too low in an attempt to reach the non-mathematical. Or it could truly be as bad as it seems.
See also: Dynamics and Cognition