[1]
Standard etymologies trace the term “level” to the balance and from there to the idea of a flat, horizontal landing, as in the stories of a building. From there, it is easy to see how the metaphor might be extended to the kinds of hierarchy discussed in this paper.
[2]
One consequence of the following discussion is that not every account of levels must offer a unique answer to the placement question. Levels of mechanisms are defined by their distinctive relata and relations; these constraints, by themselves, offer no unique answer to the placement question. This is why levels of mechanisms are, as I will argue, local rather than monolithic.
[3]
For the relevant sense of a scientific “field”, see Darden & Maul (1977) and Darden (1992).
[4]
Wimsatt’s diagram in Figure 1 reflects this. The branching tree structure is ordered by compositional relations. Wimsatt’s view of levels as dissipating waves (see Figure 3 below) flouts that relation.
[5]
One might, analogously, arrange a hierarchy of activities, with different activities occurring on different temporal scales. The idea of levels of mechanisms combines these two ideas; it is a hierarchy of doings framed by a relevance relationship between those at lower levels and those at higher levels.
[6]
It is hard to say how it would be tested and, in particular, how predictability is to be measured. Surely items in the valleys of this diagram are not unpredictable, full stop. Rather, they are more difficult to predict for creatures like us, unaided by machines and programs. It is not clear why human cognitive abilities should have any further ontological significance.
[7]
Supplemental conditions might be added to make realization more demanding (e.g., Melnyk 2003; Haug forthcoming). The point I wish to make doesn’t turn on this matter.
[8]
I discuss a representative quote from Lewis below when considering causal relations between levels of mechanisms.
[9]
One can, in cases of multiple realization, intervene into the parts and their organization without intervening to change the property of the whole, and this affords some measure of independence. Perhaps one can find room in this view for the idea of understanding bottom-up relations in a hierarchy of realization as causal (though, again, realization or token identity seem to be better ways of talking). But there is no room in the view (no conceptual room) for causation to work from the top down in such levels. For a penetrating discussion of this matter and its implications for causation in a multilevel world, see Baumgartner 2010, 2013; Romero (forthcoming).
[10]
Levels of control and levels of processing, in contrast, are defined in terms of causal relations. In levels of control, the relata are agencies and the relation is dominance. Items at higher levels direct or regulate the activities of their underlings. Majors and corporals, queen bees and drones, bosses and workers occupy different levels of a control hierarchy. Analogous relations are sometimes found among physiological systems. When one speaks of “executive function” in cognition, one is describing levels of control.
The idea of control or dominance is a causal notion, and it is independent of matters of size (witness the sauropod brain). Contra Fehr (2004), the idea that the world is organized in levels of realization or organization (as defined below) is not an expression of patriarchy; it is an equivocation to characterize realization and organization as relations of dominance. In levels of control, the relata are logically independent and spatiotemporally distinct interactors. It is not at all implausible for one to control the other causally (more on this below).
In levels of processing, the relata are processing units of some sort (such as brain regions or computational modules), and they are related as “upstream” or “downstream” in the flow of information or the order of production. In the early visual system (neglecting feedback for the moment), one can describe visual information as passing from lowest- (shallowest-, earliest-) level processing in the retina to higher- (deeper-, later-) level processing in the Lateral geniculate nucleus (LGN) and the primary visual cortex. Levels of realization and organization are not earlier or later than one another. Craik and Lockhart define levels of processing in terms of depth of semantic or cognitive processing, not in terms of decomposition.
[11]
I am here using the terms “part” and “whole” in an intuitive and inclusive way. Much of the literature on the metaphysics of parthood is simply unrelated to the many senses of part and whole used in the theories of the special sciences. I am not thinking only of objects or sets, but also about events and temporal units. I sketch a more restrictive kind of part-whole relation below, but this remains an open question (see Sanford 1993).
[12]
Elephant hearts are parts of elephants, and puffin hearts are parts of puffins, and hearts are parts of organisms. Yet puffin hearts are not at a lower part-whole level than elephants, and elephant hearts are not at a lower part-whole level than puffins.
[13]
I shall not enter here into the difficult question of how token properties, processes, and objects are individuated in the biological sciences. Consider the spatial memory system in a rat. If we take this to be one thing over the life of the organism, then it will be composed of many different sets of parts over the course of its existence, like the ship of Theseus. If we take it instead to be the spatial memory system involved in learning the layout of one particular maze, which learning might be constituted by multiple trials and extended investigation, then again we will have a single higher-level system composed differently at different time slices. If we focus on a given instant in time, then no learning can occur; learning is a kind of change. For now, I simply note (along with Bechtel & Mundale 1999) that the appropriate mapping between such parts and wholes presupposes a criterion of individuation for the whole, and what counts or does not count as a part will be determined by whether it contributes to that whole, however specified. This is related to Marcus’ (2006) thought that any token identity between levels presupposes a (non-dummy) sortal that fixes the individuation conditions of the relata in the same way. Carl Gillett (2013) has called attention to the need for different accounts of compositional relations for properties, processes, and objects. Here, I am glossing over these differences to keep the discussion simple.
[14]
Marr’s levels are not space-involving in this way. The algorithm is not located within the computation, it is not a substage of the computation, and it is not organized together with other parts in the service of the computation.
[15]
This is not a big departure from classical mereology; one could simply restrict one’s attention to proper parts. But the point underscores the fact that classical mereology was not developed with an eye toward understanding the sense in which pyramidal cells are parts of the hippocampus.
[16]
Clearly Eric’s muscles are not part of the team, but this reflects only the fact that teams can have only certain kinds of part as members. If we look rather at an activity of the whole and ask what contributes to that, a different picture emerges.
[17]
Specific details about lower-level parts might be screened off in cases of multiple realization. Specific details about the parts might not be relevant. In that case, it would appear one must appeal to more abstract properties of the parts.
[18]
Once we have made this adjustment, the relata in this relevance-mereology are no longer objects but rather properties, activities/processes, or (as is more common in philosophical parlance) events. One does not explain the elephant; one explains why the elephant has large ears or how the elephant circulates its blood. One does not explain gasses; one explains their temperature and pressure. This point marks a significant departure from the mereological views of levels discussed above. Each of those applications of the levels metaphor focuses on objects or types of objects (societies, organisms, cells, and so on) as the relata, not on their properties, activities, and aspects of their organization.
In many cases, the components picked out in a mechanistic decomposition fail to correspond to paradigmatic objects with clear spatial boundaries. The synapse, for instance, is composed of part of a presynapatic cell (the axon terminal), part of a post-synaptic cell (the dendrite or bouton), and a gap between them. What unifies these items into an object is their organized behavior: the pre-synaptic cell releases transmitters that traverse the cleft and act on the postsynaptic cell. Synapses are not cells or parts of cells, nor are they composed of cells. Rather, they are objects unified by their relevance to a given actvity of the whole, such as chemical transmission.
[19]
Levy (2013) calls attention to the fact that biological systems typically involve both aggregation and organization.
[20]
These might be understood as the obtaining of a property or the unfolding of a process over time. What counts as a static property often depends on one’s temporal resolution.
[21]
I have not always chosen my language in a way that comports with common usage among metaphysicians, preferring to follow Salmon (1984). In this paper, I have tried to make it clear that I am interested in components. Components, as the name suggests, compose behavior of the higher-level mechanism when organized together. All of the component entities and activities organized together, it now seems appropriate to say, jointly constitute the behavior of the whole. That is, I am now using componency to talk about relationships between wholes and parts, and I am using “constitution” to talk about levels of micro-realization. I am not especially interested in the relationship between statues and lumps of clay. I am interested in how parts are organized and interact so that together they exhibit higher-level behaviors. I know of no metaphysician who has developed an adequate notion to express this, so perhaps I will be forgiven for appropriating these words for new uses.
[22]
Nothing in this picture is meant to deny token identity between the behavior of a mechanism as a whole and the organized behavior of its parts. Because there are some conceptual difficulties that stand in the way of speaking meaningfully about token identities between levels, I have written with fewer commitments about constitution.
[23]
This has struck some readers as circular because it appears to state that X and S are at the same level if they are not at different levels. Appearances to the contrary, this is not circular. I have defined “same level” in terms of the notion of “different level” and the latter is defined in terms of componency relations. The appearance of circularity, I believe, results from the fact that most people assume that the notion of “same level” must be primitive relative to the notion of “different level,” and I have reversed that assumed order.
[24]
Another way to see that levels of mechanisms do not answer the placement question is to recognize an apparent. Suppose X1 and X2 are components in the same mechanism, that neither is a component in the other, and that the behavior of X2 can be decomposed into a set of interacting components, including P1. X1 would, according to this account, be “at the same level” as both X2 and P1 even though X2 and P1 are at different levels from one another. This problem, first raised by Lindley Darden (personal communication), is only a problem if one demands that there must be a unique answer to the placement question for an account of mechanistic levels. My argument against the notion of monolithic levels turns on the absence of any good principle for stretching the ideal of levels beyond its local context.
[25]
This idea of reduction is not the standard notion of theory reduction but something closer to what Eric Kandel means by the term: choosing to study complex phenomena in extremely simple systems. We might call this experimental reductionism.
[26]
Not in the Salmon (1984) sense, but in the colloquial sense of an unfolding sequence of states and activities.
[27]
This case is easiest to make for interventions that start the process midway. If an intervention, instead, were to augment C and thereby produce a more potentiated synapse than one would otherwise have had, then causal language would appear to be appropriate. The intervention changes, makes a difference to, the input-output relationship. These considerations generalize naturally to claims about types of mechanistic parts and wholes. Separately: When we describe this relation as a kind of production, levels show up as intermediate causes. Perhaps the temptation to speak of levels at all is lessened if one maintains that perspective. But this is not a change in what is being said so much as a change in how it is being said.