Monday 20 June 2016

Mechanisms and Models of Mechanisms (#MechanismWeek 1)

In this first #MechanismWeek post, I will define a mechanism and briefly describe the kind of models of mechanisms you can build. I begin with various kinds of functional models (Cummins, 1975, 2001; Weiskopf, 2011). These either break capacities of systems into more coherent, easily studied sub-capacities (think of breaking memory into long term memory and short term memory as a simple example) or model them with components that may or may not be really implemented in the organism (e.g. geonsexemplars).

I will then introduce the idea of a dynamic causal mechanistic model (Bechtel & Abrahamsen, 2010) which attempt to map model components directly onto the real parts and processes of the mechanism at hand. The argument is that while functional models provide useful descriptions of mechanisms, they do not provide an explanation of that mechanism, and that only mechanistic models can explain.

Mechanisms
Bechtel & Abrahamsen (2010) characterise a mechanism as follows:
A mechanism is a structure performing a function in virtue of its component parts, component operations, and their organization. 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 phenomena.
pg 323
This is fairly straightforward. Mechanisms are real things that contain real parts arranged and connected in particular ways so as to do a particular thing. The idea is that in science, one of our goals is to model the mechanisms that implement particular behaviours, where a model is a formal (typically mathematical) representation of the system under study.

Non-mechanistic Models of Structures
Analytic Functional Models (Cummins, 1975, 2001): Things happen out in the world because of the operation of mechanisms, but that doesn't actually mean our models have to be mechanistic. One strategy is to develop analytic functional models (Cummins, 1975, 2001). Functional analysis take a capacity of a system (the phenomena that the mechanism causes) and breaks it down into a set of smaller, connected, easier to understand capacities. These sub-capacities do not necessarily map cleanly onto any real components, but this is not the goal here.

This should actually sound very familiar to all the psychologists reading this, because it is a big part of what we do. Take the psychological capacity memory. The history of the field has been one of decomposing 'memory' into smaller, more coherent sub-capacities. The first split (Atkinson & Shiffrin, 1968) was into long term memory, short term memory and sensory memory. Long term memory has since been broken down into episodic memory, semantic memory, procedural memory, and more. Short term memory became working memory (Baddeley & Hitch, 1974) and this was decomposed into the visuo-spatial sketch pad, the auditory-phonological loop and the central executive. See Figure 1.
Figure 1. Dividing memory into sub- and sub-sub-capacities. This ain't over yet, either.
None of these capacities refer explicitly to any empirically identified real components of a biological system; they simply refer to known capacities of a biological system (remembering humans). This fact has not stopped memory research from being very productive and mapping out many interesting features of memory.

Other Functional Accounts (Weiskopf, 2011): The defining feature of the functional analysis is that they do not refer to any components, only sub-capacities. Wieskopf (2011) notes that many cognitive models do, in fact, refer to components, just not necessarily real components. He examines examples such as object recognition (e.g. Hummel & Biederman, 1992) and categorisation (e.g. Kruschke, 1992). These models do generate the main capacity with components (things like geons or exemplars) but these components do not immediately map onto anything in the person (although obviously people go looking; e.g. Bülthoff, Edelman & Tarr, 1994 hunting for exemplars in the brain). Many representational theories are this kind of model and again they have been extremely productive.

Mechanistic Models of Structures
To count as a mechanistic model (i.e. one that accurately represents the actual structure in the world producing the behaviour), all terms in the model should represent a real part or process (the 'Real Components Constraint'; Craver, 2007) and they should be organised in such a way that when you run the model, the relevant behaviour comes out. This model is a model of 'how-actually' the behaviour occurs in the system being modeled (as opposed to being simply a description of 'how-possibly' or 'how-probably' that behaviour could happen; Craver, 2007). 

In order to develop a mechanistic model, one must empirically decompose the system correctly to identify (localise) the actual parts and processes, and then build the model with reference to the results of those studies. In other words, the model is the thing you build on the basis of empirical research specifically designed to identify the real components of the mechanism under study. If you do things this way, you end up with a dynamical mechanistic explanation of the mechanism; not simply a descriptive model, but an explanatory model of the behavior of interest.

Bechtel & Abrahamsen illustrate this with a fairly deep dive into explanatory models of the mechanisms of circadian rhythms. They walk through a series of experiments on the actual biological systems implementing these rhythms that identified the relevant real components and how they were coupled together that enabled the development of mechanistic models of circadian rhythms.

There are 6 key advantages gained from genuinely dynamical mechanistic explanations, and only these models give you all 6 advantages (Bechtel & Abrahamsen, 2010):
  1. demonstrate that a given mechanism is sufficient to produce the target phenomenon
  2. explore the functioning of the mechanism in a larger parameter space than is accessible in experiments
  3. identify whether candidate parts are essential to the mechanism’s functioning
  4. explore how particular types of damage might affect the system by perturbing the model in particular ways
  5. to explain how coordinated behavior can emerge from the coupling of simpler mechanisms
  6. to explore the consequences of altering the relations between multiple mechanisms
In essence, having a mechanistic model is like having the mechanism ready to hand but in your computer, where you can poke it, prod it and do all kinds of things you can't do to the real system, and have all that work genuinely further your understanding of the real system. This seems like quite a good idea, but psychology does not yet have these (more on this in the next post).

Summary
Functional models are often a good start (for example, we now understand that the mechanisms underpinning LTM and STM are likely quite different because these sub-capacities behave so very differently). But Bechtel & Abrahamsen advocate that mechanistic, 'how-actually' models should be the goal of any science because they're just so gosh-darn useful. In particular, they are the only model that counts as an explanation of the mechanism.

In order to get such a model, you must first empirically characterise the mechanism in question, identifying the real parts and processes (the real components) that actually make the mechanism exhibit the behaviour in question. Only then do you try to model it, building a model with terms that explicitly represent all the necessary real components arranged in the way they are arranged in the mechanism - you must model both the composition and the organisation of the system at hand.

Not everyone agrees that mechanistic models are the only ones with explanatory clout. Weiskopf (2011) argues you get at least some good explanation out of functional analyses, and Silberstein & Chemero (2013) argue that dynamical models do good work too. The next two posts will cover these kinds of models in more detail.

References
Atkinson, R. C., & Shiffrin, R. M. (1968). Human memory: A proposed system and its control processes. The Psychology of Learning and Motivation, 2, 89-195.

Baddeley, A. D., & Hitch, G. (1974). Working memory. The Psychology of Learning and Motivation, 8, 47-89.

Bechtel, W., & Abrahamsen, A. (2010). Dynamic mechanistic explanation: Computational modeling of circadian rhythms as an exemplar for cognitive scienc`e. Studies in History and Philosophy of Science Part A, 41(3), 321-333.

Bulthoff, H., Edelman, S. Y., & Tarr, M. J. (1995). How are three-dimensional objects represented in the brain?. Cerebral Cortex, 5(3), 247-260.

Craver, C. F. (2007). Explaining the brain. Oxford University Press.

Cummins, R. (1975). Functional explanation. Journal of Philosophy, 72, 741-764.

Cummins, R. (2000). How does it work?" versus" what are the laws?": Two conceptions of psychological explanation. Explanation and Cognition, 117-144.


Hummel, J. E., & Biederman, I. (1992). Dynamic binding in a neural network for shape recognition. Psychological Review, 99(3), 480.

Kruschke, J. K. (1992). ALCOVE: an exemplar-based connectionist model of category learning. Psychological Review, 99(1), 22.

Weiskopf, D. A. (2011). Models and mechanisms in psychological explanation. Synthese, 183(3), 313-338.

No comments:

Post a Comment