Source
Bechtel and Abrahamsen (2010), LIT-1936, §§2, 4–5, pp. 322–323 and 328–332; read in NOTE-1492.
What was actually shown
-
The definition (p. 323). A dynamic mechanistic explanation takes “properties of certain parts or operations of the mechanism” into the model “as variables or parameters, thereby anchoring that model to the mechanistic account”.
-
The contrast (p. 329). “Circadian modelers begin with an existing mechanistic account … The characterization of the whole mechanism, in detail, is a prior task achieved by other methods.” For cognitive scientists, “the computational model, with its hypothesized representations and operations, is the only available account of the mechanism”, and “the only empirical check on this is the extent to which the model as a whole approximates the behavior of humans”.
-
The consequence, in a case (p. 330). Rumelhart and McClelland’s single network and Pinker and Prince’s dual mechanism each answered the other’s new phenomena by modification. “In the absence of this source of constraint … the ensuing conflicts are often irresolvable. The relevant models are underdetermined.”
-
The same limit, four more times (pp. 330–332).
- Beer’s walking controllers, the “less hypothetical” one, being partly grounded in neural research.
- McClelland et al.’s consolidation model.
- Hinton and Shallice’s deep dyslexia network.
- van Leeuwen’s lattice.
For each the paper says the inference to the actual mechanism is supported “only indirectly”, or holds only “if it in fact employs” the model’s organization.
-
What could have failed. A de novo model’s fit could have singled out its architecture, as when no rival fits at all. In the past-tense case rivals did fit, after modification.
What this does not say
- That anchored models are right. Anchoring makes the variables refer to something. The equations can still “misrepresent or leave out critical components” (p. 325).
- That de novo models contribute nothing. They demonstrate adequacy (C1) and explore parameters (C2), and Hinton and Shallice’s offered a new explanation of a disorder. Their standing is how-possibly.
- That fit can never decide. “Usually” is the paper’s word. A model can be ruled out, and a set of rivals can be pruned to one if the phenomena are rich enough. The paper gives no measure of when.
- A general theorem of underdetermination. That comes from the record’s own construction (CLAIM-519): over a fixed interface and horizon, a lookup table reproduces any agent’s interaction law exactly.
Connections
- CLAIM-519. The same thesis for world models: competence does not establish an internal model, and “establishing an internal mechanism needs evidence that discriminates candidate implementations”. This paper supplies the methodological precedent, and its name for the discriminating evidence: identification of parts and operations by experiment.
- CLAIM-515. Its third question, which mechanisms realize a capability, is what this paper says fit alone does not answer.
- CASE-189 and CASE-190. The record’s constructions are de novo in this sense. Their capacities and connectors are posited, and “Which terms are internal is stipulated”. They show how a maintained organization could behave, not that any actual system is one.
- CLAIM-522. Different mechanisms, internal producer and outside controller, with identical dynamics. This paper’s remedy, find the parts and intervene on them, is the one CLAIM-522 gives.