I argued specifically that information lets us propose mechanistic explanations for distributed cognitive systems. We recently came across the philosophical literature on what mechanisms are and how to make them, and it seemed immediately clear that we should be doing this (and that we already are; see below).
It turns out, though, that some of the radical camp (specifically Chemero and Silberstein) don't think we can have distributed cognition mechanisms, but that this is ok because we still get explanations out of our dynamical models.
This post will briefly review what dynamic mechanistic explanation is and why they are so useful (Bechtel & Abrahamsen, 2010). I'll briefly summarise the radical opposition to mechanisms, point out their answer doesn't work, then talk an example that shows we can have radical dynamic mechanistic explanatory models without giving anything up. The trick, as ever, will be to rely heavily on information as the component part that allows cognitive mechanisms to extend out over body and environment.
Dynamic Mechanistic Explanations
Bechtel & Abrahamsen (2010) describe two basic strategies for making models of mechanisms in science:
- Use the model making process to propose a mechanism, and test it be seeing if the model can produce the relevant phenomena
- Run experiments to independently figure out what pieces go into the actual mechanism then build a model from just those pieces in order to expand the range of science you can do on the mechanism
The problem is that you can only get dynamic mechanistic explanations out of Strategy 2. There are always multiple structures that can produce a given function (degeneracy), so getting the function out of your model is not evidence that the model's mechanism matches the real one. In order for your model to count as an explanation of the real thing, you need to be able to show that it's validly connected to the real thing, which requires Strategy 2. Bechtel & Abrahamsen advocate that cognitive science should change modelling strategies so it can get to models of real mechanisms too. So what is a mechanism? Their definition is as follows:
A mechanism is a structure performing a function in virtue of its component parts, component operations and their organisation. The orchestrated functioning of the mechanism, manifested in patterns of change over time in properties of its parts and operations, is responsible for one or more phenomenaOur models of a mechanism explain that mechanism when we can use it to tell the above story about the real thing. When we can explain mechanisms, we gain a lot - we can do experiments on the model, maybe ones you can't run on the system for some reason. We can use the model to expand our understanding of the real thing. Currently, cognitive science can't do this with our models.
The Radical ObjectionIn order to figure out the mechanism before modelling it, you have to experimentally break the mechanism into it's working parts (decomposition) and figure out which operations each part implements (localisation). Silberstein & Chemero (2013) note you can't do this to systems in which the parts are linked nonlinearly because the operations of the parts depends on the operation of other parts, and so studying the parts in isolation doesn't tell you how they work in the system. They propose that dynamical models are then required (which generate behavior by setting up lawful dynamics and initial conditions and running the model). Bechtel and other mechanists do not think these provide explanations; Silberstein & Chemero think they do, and talk a little about systems neuroscience and network analyses of the brain as examples.
An analysisSilberstein and Chemero are wrong about dynamical models. These do not explain. I (sort of) wrote about this 4 years ago. I was coming at it from a slightly different angle then, but ended up in the right place: purely dynamical models are no good because they don't contain anything representing the actual mechanism in them. Dynamics is the right formal language, but you need to use it in service of a theory of the actual behavior (e.g. ecological psychology).
What I had figured out but hadn't quite articulated was the Bechtel analysis: these models are useful descriptions, but they don't explain the real thing. Chemero has made this mistake before, when he used the HKB model in his book as an explanation of coordinated rhythmic movement. The HKB model is a dynamical model, but it does not contain anything that resembles or represents any actual element of coordinated rhythmic movement. It was a useful model. But crucially, it led to numerous incorrect predictions and failures (see this post and the various comments) and those were because it isn't a mechanistic model. Network analyses in neuroscience have the same problem; it's not yet clear that their properties map onto any actual parts of the brain and so while cool, they do not explain the behavior of that particular system.
So, nonlinearity can interfere with our ability to get mechanistic and dynamical models are not explanations. Are we doomed?
Dynamic Mechanistic Explanations in Radical Embodied Cognitive Science
The reason I liked the Bechtel & Abrahamsen analysis about modelling strategies is that it clearly articulated a problem I've had about models in cognitive science for a long time, and because the mechanistic strategy they advocate is exactly the strategy that lead Geoff Bingham to develop his model, which explains why it has been so successful and useful.In my talk, I laid out the ecological model of coordination as an exemplar of a dynamic mechanistic model capable of explanation and all the other things these models can do. The thing that makes the model work is including a term representing the actual ecological information used to perceive relative phase, as well as terms representing the actual dynamical organisation of rhythmically moving limbs. These terms were indeed built on the basis of previous work probing the real system, and the net result is an explanatory model we can use to extend our scientific investigation of coordination.
There is nonlinearity in the model; but crucially that nonlinearity lives inside 'components' of the model and not in the couplings. I don't think this is an accident; we've never decomposed limb movement past the level of 'nonlinear oscillator' because it makes no sense to. That is the unit and it's internal composition and organisation is defined by the behavior of the unit. The need to decompose a mechanism in order to model it doesn't require you to tunnel all the way down to the bottom, just down to the level where there are identifiable components. The radical objection fades away and we can get explanatory models into our radical science. This, I contend, is a great thing and can only help the field.
Summary
The best way to model a mechanism is to decompose the real thing and empirically identify which parts do what work (localisation), and only then formally describe that mechanism with a model made up of only those parts that do work in the real thing. We can do this for radical, ecological, enactivist explanations of behaviour, and so long as we remember one of the key components is ecological information those mechanistic models can reach out of the body and into the world. These models can then provide dynamic mechanistic explanations of behavior and our science is all the better for it.ReferencesBechtel, W., & Abrahamsen, A. (2010). Dynamic mechanistic explanation: Computational modeling of circadian rhythms as an exemplar for cognitive science. Studies in History and Philosophy of Science Part A, 41(3), 321-333. Download