Wednesday, 22 April 2015

Function vs Structure (A Purple Peril)

One of the apparently controversial things that I say is that psychology, as a science, needs to address function before it gets worried about structure; what is the brain trying to do, vs how is it doing it? This Peril lays out the argument in a little more detail. As always, this is my current thinking not my final thinking and I am happy as ever to hear arguments for and against this proposal.

Structure (the details of how a function is implemented) is important, there is no doubt. But I see two related arguments about putting function first, or at least giving it the driver's seat in our science.

Argument 1: Function provides the job description required to constrain our understanding of brain activity
The first argument in favour of function is that cognitive neuroscience is essentially in the business of uncovering the neural implementations of the things cognitive science has identified need doing. Cognitive science provides the 'job description' for the brain; the functional analysis of what the brain is up to that guides the interpretation of the structural analysis that comes out of neuroscience. See this new paper in Cognition making this argument in more detail. Without this theory, all neuroscience can tell you is that bit is currently talking to that bit. With the theory, it can talk about why they are talking to one another and what thtat talk is achieving.

My problem, of course, is that I think radical embodied cognition is working to overturn the current functional analysis that guides modern neuroscience. If REC is right, then there are implications for how we do neuroscience, and I worry that a lot of neuroscience might go away; who needs to know where in the brain object representations live if there are no such things? I therefore think we need to get the functional story right first.

Argument 2: Biological systems work to preserve function, not structure
The second argument in favour of function is that the evidence strongly suggests that perception-action and neural systems work to maintain function and not structure. That is, in a very real sense, these systems don't care what they have to do or what shape they have to take so long as they fulfill the relevant task demand. This supports the idea in Argument 1 that it's the functional analysis that defines what the system will look like.

Perception-action: When you reach to grasp an object you don't try to maintain a particular form of the movement, you tailor the spatial and temporal characteristics of the movement to the current situation. We can do this because perception-action systems typically have access to more degrees of freedom (things that can change state) than required to solve the task. This fact is typically presented as a control problem (anything that can change state must be controlled and that control gets very hard, very fast as you increase the number of things that can change). It can also be framed as a feature, not a bug; the 'bliss of motor abundance', in Latash's excellent phrase. Our ability to reconfigure our actions to account for local conditions is the thing that enables our successful, functional interactions with the environment. 

Neural systems: The brain has preferred ways of doing things, as does the perception-action system. But it has remarkable resilience in the face of damage, disease, atypical development and, presumably, the day-to-day variation in the conditions under which you have to exhibit the "same" behaviours. In neural systems this gets referred to as degeneracy; the ability of two networks with different configurations ('wiring') to produce the same function. It is again a feature, and not a bug. My favourite example of this is Eve Marder's work demonstrating how there are 452,516 biologically plausible different ways to hook up the neurons in a lobster gut and still have it output the critical pyloric rhythm that moves food through that gut (Prinz et al, 2004blogged in detail here). 

Can't we just do them both in parallel?
This is the most common question I get about this topic; can't we pursue functional questions and structural questions at the same time? My answer is sure, you can, but it's risky. If you get the functional analysis wrong, then you will have spent a bunch of time asking pointless questions about the structure of the nervous system. (We think this has already happened.)

Surely the neuroscience data will tell us which way to jump, function wise? Well, partly. But there is no such thing as theory free data. Your guiding theory (the functional analysis) dictates what questions you ask and which ones you don't ask, and how to interpret the results. This is ok, it's what theory is for. But that means that structural data collected under the assumptions of one functional analysis can't tell us much about how useful the other functional analysis might be. Until neuroscience collects some data under the assumptions of REC, we don't have the ability to compare and contrast the results (anyone who knows of useful work, please tell us about it in the comments!)

(Of course, evidence (from neuroscience and elsewhere) can feed back into the functional analysis. A famous example is Freeman and Skarda (1990) who detail how they spent a long time trying to understand olfaction in the brain from a representational stance and discovered that their data simply did not support that kind of functional analysis. So in principle, these two threads can inform one another. But in practice, my suspicion is that the functional analysis dominates for the reasons laid out above). 

Further Reading
Freeman, W. J., & Skarda, C. A. (1990). Representations: Who needs them. Brain organization and memory: Cells, systems, and circuits, 375-380. Download

Prinz, A., Bucher, D., & Marder, E. (2004). Similar network activity from disparate circuit parameters Nature Neuroscience, 7 (12), 1345-1352. Download

Related Blog Posts
There's More Than One Way to Rhythmically Move a Lobster


  1. Here's a systems theoretic view of what work needs to be done before one has licence to use the "function" word:

  2. Surely you don't mean to imply that neuroscientists have wasted their time in the discovery of the neuron, action potential, synaptic transmission, neurotransmitters, connectivity patterns, spike timing dependent plasticity, the laminar structure of cortex, etc etc etc. Otherwise, we would have to wait indefinitely while we sort out our functional analyses before starting research on the brain. Your therefore disagree with yourself: research on structure and function *in parallel* is the best approach to a complete understanding.

    1. Almost nothing you mention is about the implementation of any specific cognitive process, so they aren't what I'm talking with regards to 'structure'.

  3. But this is the wetware that are the building blocks of the implementation. They are part of the structure. E.g., independent of cog sci, it has been noticed that connections between brain areas tends to be massively reciprocal. In fact, feedback connections sometimes outnumber feedforward connections. This architecture constrains cognitive models.

    1. Yes, but those basic mechanisms might be computing or they might be being nonlinear dynamical systems or they might do something else. This structure by itself does not uniquely lock down the function of that structure.

    2. The structure surely provides clues, though.

    3. Hi Andrew,

      Thanks for the blog and the good conversations. Most of it is over my head but here are two questions:

      1 - Could those "basic mechanisms" be computing AND being parts of nonlinear dynamic systems?

      2 - Even if the structures do not "lock down" function, do they constrain the possibilities for function?

    4. Andrew will probably convey this much more eloquently and concisely, but: Tim van Gelder addresses this in his paper The Dynamical Hypothesis in Cognitive Science (full text: In fact, this is a common objection to the Dynamical Hypothesis, that computation is somehow needed. I would refer to sections 6.3, 6.4, and 6.9 in which van Gelder replies to objections that dynamical systems are computers or computable (6.3 and 6.4) or that an understanding of structure requires a computational approach (6.9).

      I think that Computationalists and Dynamicists are interested in different things (this doesn't necessarily mean they are mutually exclusive, though). Dynamics is all about change over time, whereas the former abstracts away from that, to some extent. Dynamicists talk about states geometrically, in terms of position w/r/t other states of the system, whereas Computionalists tend to focus more on the internal combinations of syntactic structure. (I hope I am not misconstruing anything, it's not my intention).

    5. Greg, I obviously can't speak for Andrew, but... to address your concern... I find that I almost never have a problem talking with neuroscientists, while I frequently run into difficulty talking with cognitive neuroscientists. That is, I don't think anyone in this discussion has a problem with research into the biological structure of the brain, and, to the contrary, are fascinated by it. Also, we tend to get a long with brain people.

      The question is whether we have jumped ahead of ourselves in the pursuit of "cognitive neuroscience."

      Those trying to merge their perfectly good neuroscience with a faulty theory of psychology will run into predictable problems. If we are right, billions of research dollars have been spent in the pursuit of the neural equivalent of aether or phlogiston. While that research won't exactly amount to nothing in the future, it will amount to very little in light of the money invested. Given that, for the past 20 years or so, results in cognitive neuroscience have been generally considered not-worth-citing 2-4 years after publication, even the tippy top journals, it is not an absurd position.

  4. For those with an interest in this kind of thing: the Walter J. Freeman who co-authored the non-representational neuroscience article (with Skarda) linked to above is the son of the infamous Walter J. Freeman II who performed thousands of so-called "icepick" lobotomies.

  5. It seems to me that while you may have a fairly good definition of structure, you need to define function better. Because different nexi of the structure will also have a structural function (spandrel like) - which is often neglected.

    But the biggest problem is that neuroscientists often talk about function with structure in mind (e.g. when conflating collocating functions in the brain with other functions and drawing causal connections) and, of course, it's hard to even imagine how the structure would be studied without some function in mind. Even at the level of the neuron people refer to connections in a rather anthropomorphic way.

    And of course, the so called 'functions' will also have some built in structures that may get overlooked as such. This leads to what I've called the 'isomorphism fallacy' - the assumption that the structures one idenfies in 'cognition' (most broadly conceived from perception to language or social cognition) must have neural correlates that can map onto them - the hidden assumption behind phrenology and localism. Or, going the other way, that now that we know the brain uses connected networks, we must look for hitherto unnoticed connected networks in thought, language, perception, etc. (

    Perhaps a better way would be to come up with a construction-grammar-like approach to always consider function-structure pairings at the level where structure and function can reasonably be seen as paired rather than imagine two floating but somehow interconnected worlds of structure and function. This is how I am interpreting you embodiment approach and why I think it's so important. But I feel it's key not to try to overgeneralize it out of all context.

  6. I wonder how this relates to this science of psychiatry and even anti-psychiatry? Just wondering?!