One of the dominant frameworks of explanatory practice in the neurosciences and the biological sciences in general is the model of mechanistic explanation proposed in its modern form by Bechtel & Richardson (1993) and recently extended by Carl Craver (2007). Mechanistic explanations describe entities and activities that together bring about a phenomenon of interest (Machamer et al. 2000). When we are interested in how vision works, for example, we try to localize the relevant parts of the brain, and identify components and their types of interactions in order to understand how we can see things (Bechtel 2008). This model of mechanistic explanation is thought to capture the dominant explanatory practice in the biological sciences (Bechtel & Richardson 1993), but normative claims are also made with respect to the adequacy of explanatory accounts. Craver (2007) proposes a number of constraints on constitutive mechanistic explanation in order to decide whether a mechanistic model is viable or not.
In his target article, Michael Anderson (this collection) takes current models of mechanistic explanation as a starting point for proposing an important extension of the existing accounts. In previous models, the system that exhibits a phenomenon and the mechanism that explains the phenomenon were not separated. Sometimes parts of the system can be screened off with respect to the phenomenon at hand. The windshields of a car and its radio components are not really important in order to understand how it drives, for example. It’s fine to say that the whole car drives, but that only the relevant components (engine, axles, tires) are doing the mechanistic work. Focusing on the essential components of a mechanism within a larger system is unproblematic. But Anderson worries about more complex cases in the neurosciences where the system displaying a phenomenon does not encompass the relevant mechanism producing the phenomenon and might not even be on the same level of description as the mechanistic components.
Anderson wants to demonstrate that componential constitution is not sufficient as a model of mechanistic explanation for the processing of directional selectivity in the retina. Mechanisms computing direction of motion are already available at the earliest stages of the visual hierarchy. The vital components of direction selectivity in the retina could be identified. In particular, in recent discussion starburst amacrine cells (SAC) have been viewed as a mechanistic substrate of motion processing. The SACs receive input from bipolar cells, which are not themselves directionally selective, and provide output to direction-selective ganglion cells (Zhou & Lee 2008). The SACs themselves seem to be the core component for retinal motion selectivity (Park et al. 2014; Yoshida et al. 2001).
Examining the current models of how direction selectivity is created in SACs, Anderson takes note of a discrepancy between how direction selectivity is mechanistically achieved and to which parts it is ascribed. He argues for a distinction between the system S that Ψs (that is, exhibits direction selectivity) and the mechanism M that accounts for S’s Ψ-ing. For the case at hand, the SACs themselves or even just single dendritic compartments of SACs Ψ, but a much broader network of neighboring SACs and bipolar cells needs to be considered in order to provide a mechanistic account of SAC direction selectivity. Anderson proposes this distinction as an important extension of Craver and Bechtel’s model of mechanistic explanation. This has two major advantages, according to Anderson: (1) there can be entities and actions that play a role for M, but are not necessarily parts of S. This allows a certain flexibility in defining the system that displays Ψ, while at the same time including all relevant components in the mechanistic account of S’s Ψ-ing. (2) But if there are parts of M that don’t need to be spatially subsumed under S, neither do they need to be at a lower level than S. So even the requirement of componential constitution might be relaxed to allow for higher-level mechanistic components that play an important role in S’s Ψ-ing.
As an alternative account of the relationship between mechanisms M and the respective systems S, Anderson proposes that M acts as an enabling constraint on S:
[A]n enabling constraint is a relationship between entities and/or mechanisms at a particular level of description and a functional system at the same or a different level, such that the entities/mechanisms bias (i.e., change the relative probabilities of) the outcomes of processing by the system. (this collection, p. 12)
In the case of retinal direction selectivity, the mechanistic interaction between neighboring SACs and BCs acts as an enabling constraint for the direction selectivity of a specific SAC dendritic compartment (i.e., the system).
The most straightforward move by proponents of existing models of mechanistic explanation, as Anderson (this collection) also notes, would be to claim that the differentiation of system and mechanism is vacuous. Only the mechanism as a whole can do the work. Even in complex cases, one just has to pick out the right subparts of the network (specific synapses, specific compartments of neurons) that together produce the phenomenon of interest. Anderson provides a number of arguments against this way of extending the concept of mechanism/system, which I would like to briefly summarize:
Neuroscientists just don’t talk about complex directionally selective networks, but about the direction selectivity of certain dendritic branches.
The mechanism as a whole does not display a specific direction selectivity (it is not rightward-selective etc.), it only contributes to the specific selectivity in the respective SAC dendrites. The mechanism contributes to different kinds of selectivities in different dendrites.
Making fine-grained distinctions between subparts (synapses, axon branches, dendrites etc.) of the very same neurons that contribute to different directional selectivities is implausible.
When the whole network is said to be direction-selective (i.e., it Ψs), what about the dendrite itself? Is it supposed to only signal direction selectivity (signal Ψ-ing)? It is unlikely that a clear distinction between Ψ-ing and signaling Ψ-ing can be made.
The aim of this commentary is twofold. First, I would like to argue that the described cases can be handled by current models of mechanistic explanation when one considers the options of reconstituting the phenomena and top-down causation. Second, using another example of research on SACs, I would like to show that the straightforward ascription of direction selectivity to the SAC dendrites is at least debatable. When looking at how empirical results are often integrated with computational models of direction selectivity, it becomes clear that those phenomena can only be understood by considering the distributed nature of the involved networks.