6 Parts and wholes

6.3 Levels of organization: Aggregates and mechanisms

So let us focus on an application of the levels metaphor that is a part-whole relation and a (constitutive) relevance relation. I will not dwell here on the appropriate notion of relevance (see Craver 2007, see also Harinen 2014). For now, we can work with the idea that each part in such a hierarchy (in addition to being spatially and/or temporally contained within the whole) plays a necessary but insufficient role within a collection of parts that are jointly sufficient (but possibly redundant) for a given explanandum phenomenon (Couch 2011). That is, relevant parts might usefully be thought of as constitutive INUS (insufficient, but necessary part of an unnecessary but sufficient condition) for the behavior of the mechanism as a whole (Mackie 1973).[18]

Again following Wimsatt (1997), we can distinguish two ways that spatiotemporal parts contribute to a property or activity of a whole: aggregation and organization. An aggregate property is literally a sum of the properties or activities of the parts. The current flowing through an ion channel, for example, is a sum of the charges carried by individual ions. The concentration of a volume of a fluid is a sum of the number of particles in that unit volume. Aggregative properties change linearly with the addition and removal of parts. And aggregative properties do not change as the parts are intersubstituted with one another. Some properties of the hippocampus, such as its mass, remain the same when the cellular constituents of the hippocampus are reorganized to represent His Holiness. Other properties of the hippocampus, such as its information processing capacities, are destroyed. For truly aggregative properties, spatial, temporal, and causal organization among the components is irrelevant.

Aggregates are rare. The masses of the individual grains in a sand pile do, in fact, depend on the spatial distribution of the other grains (if one takes relativity seriously). What is presumed to be a homogeneous concentration of a liquid can in fact have local concentration differences depending on how the ions are organized in different parts of the fluid. In the case of non-aggregates, the activity or property of the whole is not a simple sum of the properties of the individuals. Adding or removing parts (e.g., the human heart) can lead to dramatic changes in how the system (e.g., the body) works. And rearranging the parts and their activities in space and time can eliminate the explanandum phenomenon entirely (as would happen if one randomly swapped parts of the circulatory system for one another). This is all true because spatial, temporal, and causal organization are relevant to (make a difference to, partly constitute) the property of the whole.

I use the term “mechanism” permissively to describe non-aggregative compositional systems in which the parts interact and collectively realize the behavior or property of the whole. Mechanisms are by definition more than the sums of their parts: they have properties their parts do not have, and they engage in activities that their parts cannot accomplish on their own.

Most mechanisms with which I am familiar involve myriad part-whole relations, some of which are more aggregative in nature, and some of which are less so. Many things brains do, for example, involve the flux of ions across a membrane, which flux is closer to the aggregative than the mechanistic end of the organizational spectrum. Other things brains do (such as the developing grid cells in the entorhinal cortex) require precisely organized relations among the activities of cells in and around the entorhinal cortex. This organizational spectrum from aggregate to mechanism covers all the relations that go into levels of organization, the superordinate class.[19]

In levels of mechanisms, the relata are some activity or property of a mechanism as a whole,[20] and the activities, properties, or organizational features of its components (its relevant parts and organization). Some component, X’s φ-ing, is at a lower mechanistic level than S’s ψ-ing if and only if X’s φ-ing is a component in S’s ψ-ing, that is, if and only if X’s φ-ing is a relevant spatiotemporal part of S’s ψ-ing. In levels of mechanisms (as opposed to aggregates) lower-level components are organized together to make up some behavior or property of the whole; in aggregates, the properties of the parts are summed.

Levels of mechanisms are represented in Figure 4. At the top is the activity of some mechanism as a whole (S’s ψ-ing). S’s ψ-ing is a behaving mechanism. Although one can speak of the mechanism and its activity separately (as when a mechanism stands inactive but ready to act), such separation in thought is artificial. Even the static mechanism is defined and sub-divided by reference to what it does. ψ is the topping-off activity of the mechanism for which all lower-level components are relevant. It can be idealized as an input-ouput relation, though this is an impoverished way of understanding phenomena (see Craver 2007). One level down are the activities and components, the X’s φ-ing, which compose and are organized together to constitute S’s ψ-ing.[21] Below that is another iteration of levels: the ρ-ings of Ps organized such that one of the Xs φs as it does. By organization, I mean that the parts have spatial (e.g., location, size, shape, and motion), temporal (e.g., order, rate, and duration), and active (e.g., feedback or other motifs of organization; see Levy & Bechtel 2014) relations with one another by which they work together to do something they cannot do on their own. As noted above, the relationship between levels is a part-whole relationship filtered further by constitutive relevance (Craver 2005; Harinen forthcoming). In levels of mechanisms, parts are made into higher-level components by being organized spatially, temporally, and actively into something. In more aggregate compositional relationships, they are summed into higher levels.

Figure 4: An abstract diagram of levels of mechanisms.

Contemporary theories of learning and memory provide a compelling example of levels of mechanisms (see Craver & Darden 2001; Craver 2002). The top level is a mechanism as a whole engaged in a spatial memory task, such as learning to run efficiently through a maze. One component in that mechanism, and so one level down in this description, is the hippocampus, a region of the brain thought to form a “map” of locations and orientations within the maze. The capacity of the hippocampus to acquire such an internal map of local spaces is thought to be explained, in part, by changes in synapses between pyramidal cells, specifically by a process known as Long-Term Potentiation (LTP). And it is now known that n-methyl d-aspartate (NMDA) receptors (n-methyl d-aspartate is a pharmacological agonist that binds these receptors preferentially), contribute to LTP. This story could continue downward, looking into aspects of protein chemistry and the structural changes thought to underlie channel functioning.

6.3.1 Levels of mechanisms are local

Levels of mechanisms are of entirely local significance. The levels in our example are defined by reference to a topping-off point, spatial memory, contribution to which determines whether or not a spatiotemporal part of the system is in fact relevant—whether it is a component in the mechanism for S’s ψ-ing. The hierarchy in Figure 4 and the levels of spatial memory as I have described them follow only a single (local) strand of embedding: from the behavior of the mechanism as a whole, to the behaviors of its components, on to the behaviors of one of these components, and so on.

Levels of mechanisms, like part-whole levels generally, are not monolithic divisions in the furniture of the world. Levels of mechanisms are defined only within a given part-whole hierarchy. There are different levels of mechanisms in the spatial memory system, in the circulatory system, in the osmoregulatory system, and in the visual system; the levels in each need not map onto one another. How many levels there are, and which levels are included, must be determined on a case-by-case basis by discovering which sorts of components are explanatorily relevant for a given phenomenon. Levels of mechanisms cannot be read off a menu of levels in advance.

If we apply the levels metaphor only locally, then it makes no sense to ask whether the hippocampus is at a higher or lower level than the nephra in the kidney. The nephra are not part of the hippocampus, and they are not relevant to the functioning of the hippocampus. Neither similarities of size nor similarities in kinds of part are definitive of levels of mechanisms. Rather, levels of mechanisms are defined relative to one another within a hierarchically organized mechanism.

The idea that levels of mechanisms retain some hint of the layer-cake model can sneak its way back into one’s application of the metaphor if one slides unknowingly between tokens and types of parts and wholes. Compare the following three sentences:

1. This pyramidal cell is at a lower level of mechanisms than this hippocampus.

2. Pyramidal cells are at a lower level of mechanisms than hippocampi.

3. Cells are at a lower level of mechanisms than organs.

Statement (a) expresses a mechanistic notion of levels: a particular pyramidal cell is a component of a particular hippocampal mechanism. This statement is true if the cell is a component in a mechanism for a given activity in which the hippocampus is engaged. It might be, for example, that a given pyramidal cell is a component in some hippocampal mechanisms but not others; if so, it is at a lower level to some hippocampal activities and not others.

Wimsatt describes the compositional relationship between levels as a relationship between types. He writes: “Intuitively, one thing is at a higher level than something else if things of the first type are composed of things of the second type” (Wimsatt 1976, p. 215). This is a departure from the idea of levels of mechanisms and one that threatens to reinstate something like the Oppenheim and Putnam hierarchy. Pyramidal cells are found outside the hippocampus, and those pyramidal cells are not parts in hippocampal mechanisms; they are not at a lower level of mechansitic organization. Likewise, both the hippocampus and the kidney are composed of cells; organs tend to be composed of cells. But the cells in the hippocampus are not at a lower mechanistic level than kidneys because they do not contribute to kidney function. The slide from sentences such as (a) to sentences such as (b) and (c) is a slide back toward the layer-cake model. Of course, scientists typically deal with types. But as I suggest above, this is a generalization over a relationship between tokens. The correct generalization is that the cells that compose hippocampuses are at a lower level than the hippocampuses they compose.

This is significant for two reasons. First, it helps to show that many objections to thinking about neuroscience and other special sciences in terms of levels simply do not apply to this restricted application of the metaphor. If one thinks, with Wimsatt, me, and probably Oppenheim and Putnam, that the Oppenheim and Putnam layer cake is an overly simplistic representation of the diverse ontological structures one finds in the special sciences—that things like occular dominance columns and synapses don’t readily fit that picture—one can nonetheless retain the idea that mechanisms are susceptible to multiple nested decompositions. These are different applications of the level metaphor. Secondly, and perhaps more importantly, the idea that levels are local significantly shifts the reductionist world-view for which Oppenheim and Putnam developed their ontology of levels. If one thinks of levels as levels of organization (as levels of mechanism and levels of aggregation), then it is inaccurate to think of reduction as involving relationships among theories developed to describe the items at a particular monolithic level. If reduction is simply a matter of explaining a higher-level phenomenon in terms of the organized activities of components, then reduction is still possible within a mechanistic world-picture, but it will be achieved not through grand reductions of overarching theories, but rather through piecemeal explanatory achievements for specific phenomena. Visions of the unity of science through interlevel reduction have to be revisioned not as grand unifications across the whole of science but rather as local explanatory successes. Such local explanations will, in fact, integrate findings from different sciences and bring different theoretical vocabularies into conversation with one another (see Craver 2005; Craver & Darden 2001), but it only deceptively resembles the layer cake that Oppenheim and Putnam sketched as a working hypothesis.[22]

6.3.2 Placement is weak and derivative in levels of mechanisms

One consequence of the mechanistic application of the levels metaphor is that there is no unique answer to the question of when two items are at the same mechanistic level. Only a partial answer is avialable: X’s φ-ing and S’s ψ-ing are at the same level of mechanisms only if X’s φ-ing and S’s ψ-ing are components in the same mechanism, X’s φ-ing is not a component in S’s ψ-ing, and S’s ψ-ing is not a component in X’s φ-ing.[23] Unlike size levels or levels defined in terms of the types of objects found at a given level, levels of mechanisms are defined fundamentally by the relations question: by the componency relationship between things at higher and lower levels. If two things are not related as part to whole, they are not at different levels, and so, if they are in the same mechanism, they are, in this very weak sense, at the same level. But this is just to say that sameness of level has no significance within this application of the metaphor.[24]

If one thinks of levels of organization as levels of aggregates and levels of mechanisms, then spatial containment and size relations between levels follow as an accidental consequence of the componency relation. The pyramidal cells are contained within the hippocampus, which is contained within the spatial memory system. The activities of these entities are also related as temporal part to whole: the binding of glutamate is a temporal component in the activity of the NMDA receptor. The objects at lower levels are smaller than (or at least no larger than) the whole, giving the hierarchy a derivative size ordering. Relations of size, rather than defining what it is for an item to be at a level (the placement question), are derivative from the more fundamental relationship between levels (the relations question): namely, the relationship between a mechanism and a component.

6.3.3 Emergence and levels of mechanisms

Mechanisms do things that their components taken individually cannot. This marks a sharp distinction between levels of mechanisms and levels of realization. Kim says this point is “obvious but important”:

This table has a mass of ten kilograms, and this property, that of having a mass of ten kilograms, represents a well-defined set of causal powers. But no micro-constituent of this table, none of its proper parts, has this property or the causal powers it represents. H2O molecules have causal powers that no oxygen and hydrogen atoms have. A neural assembly consisting of many thousands of neurons will have properties whose causal powers go beyond the causal powers of the properties of its constituent neurons, or subassemblies, and human beings have causal powers that none of our individual organs have. Clearly then macroproperties can, and in general do, have their own causal powers, powers that go beyond the causal powers of their micro-constituents. (Kim 1998, p. 85)

Through aggregation or organization, wholes have causal powers that their parts individually do not have. An activity at a higher level of mechanistic organization is quite literally more than the sum of its parts. It is not an aggregate. The addition and removal of parts leads to nonlinear changes in the behavior of the mechanism as a whole. It matters how the parts are organized; it is in virtue of their organization that they have properties that go beyond the properties of the individual parts (Wimsatt 1996). This feature of levels of mechanisms is so obvious, so prosaic, and so banal as to be hardly worth mentioning. No fancy complexity is required: two toothpicks stacked perpendicular to one another have the mechanistically emergent capacity to act as a lever or catapult; neither toothpick can do so on its own.

Of course, most mechanisms in biology are substantially more complicated than that. They have many more parts. Those parts interact with one another with bewildering complexity. Often they contain feedback relations that introduce nonlinear interactions into the operation of the mechanism itself. The mechanisms of LTP, for example, have yet to yield their secrets completely despite the dedicated attention of thousands of researchers over forty-odd years. A glance at any recent textbook is enough to convince one that LTP involves myriad intracellular reactions, protein synthesis, structural features of dendritic spines, changes to vesicular release, and retrograde transmission with nitric oxide. The mechanism involves so many parts and interactions that it would be useless, if not impossible, to represent them all in a visual diagram. Keeping track of how they all work together would require a very complicated computational model of some sort that has yet to be developed. As mechanisms get this complicated, we reach the limits of our ability to predict how the behavior of the whole will change as the parts change. Any change introduced to a part has so many ramifications that it is difficult or impossible for creatures like us to keep track of them all. This is an interesting fact about us and the limits of our cognitive and modeling prowess. But, ontologically, it is the same old banal fact about the importance of organization in mechanisms. We have added only that we have difficulty keeping track of it all.

Likewise, a common scientific complaint against reductionistic research programs in biology and neuroscience is that one can make only limited progress by studying the parts of mechanisms in isolation from one another.[25] We can study LTP in cells grown in a culture, and we can study hippocampal computations in a razor-thin hippocampal slice, and we can study spatial learning in highly contrived environmental settings such as a large pool filled with milky water (the Morris water maze). Such reductionist practices are absolutely essential to progress in the sciences. Nonetheless, one engaged in such practices must (and typically does) bear in mind that the behavior of the part when it is isolated for experimental purposes might be very different from the behavior the part exhibits when it is working in the context of a mechanism. Causal interactions with other parts of the mechanism and background conditions “in the wild” might lead to behaviors that would never be discovered in such simplified preparations. This is an extremely important point about reductionist research programs (Bechtel & Richardson 1993), and one might choose to describe this well-known difficulty with the language of emergence. But this is just to say that one cannot truly understand how a mechanism works until one understands how all its parts are organized together and working in the relevant conditions, and this we have already said repeatedly.

I emphasize the banality of these observations to stress that many of the things one wants to say about organization in biological systems can be said within the mechanistic application of the levels metaphor without introducing anything that is metaphysically novel or suspect. As the complexity of a mechanism increases, the epistemic challenges we face in discovering and modeling it increase as well, but this is of no significance for the ontic structures—the entities, activities, and organizational features that exist in the world.

Not so for spooky emergence. Spooky emergence is spooky precisely because it is commited to the existence of higher-level properties that have no explanation in terms of the parts, activities, and organizational features of the system in the relevant conditions. Levels of mechanisms are levels of ontic mechanistic explanation (Craver 2014): they are defined in terms of componency and constitutive explanatory relevance. If that explanatory relationship is severed, then the sense in which emergent properties are at a “higher level” must be altogether different than the compositional notion of levels in levels of mechanisms. If one imagines that atoms compose molecules, which are organized into cells, which are linked into networks from which mental properties spookily emerge, the first three steps are upward steps in a hierarchy of levels of mechanisms, but the last is not. The ability of organization to elicit novel causal powers (that is, nonaggregative behaviors and properties) is unmysterious both in scientific common sense and common sense proper (Van Gulick 1993; Kim 1998). Appeal to strong or spooky emergence, on the other hand, justifiably arouses suspicion. Indeed, it is unclear why properties that emerge in a spooky fashion should be thought of as higher-level at all. Perhaps the very idea of spooky emergence is incoherent.

6.3.4 Mechanistic levels are not causally related to one another

As with levels of realization, many common assumptions about the nature of causation would appear to make causal relations between mechanistic parts and the properties or behaviors of wholes suspect. Items at different levels of aggregation and at different levels of mechanisms are intimate with one another in much the same way that items at different levels of realization are intimate. Lewis is explicit. If C causes E:

C and E must be distinct events [if they are to be causally related]—and distinct not only in the sense of nonidentity but also in the sense of nonoverlap and non-implication. It won’t do to say that my speaking this sentence causes my speaking this sentence; or that my speaking the whole of it causes my speaking the first half of it; or that my speaking causes my speaking it loudly, or vice versa. (Lewis 2000, p. 78)

The relevant kind of intimacy for levels of mechanisms is overlap between token events or processes. The relationship between LTP and the opening of NMDA receptors during LTP induction is directly analogous to the relationship between speaking the whole of a sentence and speaking its first half. The induction of LTP is partly constituted by the opening of the NMDA receptor. The would-be cause in this top-down causal claim already contains the would-be effect within it. There is nothing additional to be produced in the effect; the occurrence of the effect includes the occurrence of the cause.

What about the bottom-up case? We might say that the spark plugs cause the engine to run, all the while acknowledging that the sparking of spark plugs is part of the operation of the engine. The naturalness of this locution is at least partly due to an ambiguity in the way we commonly describe the behavior of a mechanism as a whole. Sometimes we describe it as an activity or process that starts with the mechanism’s setup conditions and ends with its termination conditions (Machamer et al. 2000). Thus we might describe Long-Term Potentiation as a process[26] or activity beginning with a rapid and repeated stimulus (called a tetanus) to the presynaptic neuron and ending with enhanced transmission across the synapse. Other times we describe the behavior of the mechanism as a whole, the phenomenon, as the product of that process (or one of its termination conditions). Thus we might say that the mechanism of Long-Term Potentiation produces a potentiated synapse. This way of speaking leads us to think in terms of the antecedent causes of potentiation: the tetanus is a distal cause, and the subsequent changes in the NMDA receptor are more proximal. If we think about the behavior of the mechanism in the second way, as a product, it is natural to think of the opening of the NMDA receptor as a cause of the synapses being potentiated (and indeed it is). But if we think about the behavior of the mechanism in the first way, as an input-output relation starting with the tetanus and ending with a potentiated synapse, then it is wrong to think of the tetanus or the opening of the NMDA receptor as a cause of that. The NMDA receptor is a part of that causal process. These are two equally acceptable ways of describing the relationship between a mechanism and a phenomenon; they are easily translated into one another. However, if one is careless, these ways of speaking and writing invite equivocation of precisely the sort that we are struggling here to avoid.

Figure 5: Why bottom-up causation is conceptually problematic.

Suppose we represent the input-output relationship constituting LTP as in the top of Figure 5, where I is a tetanus and O is a stable, potentiated synapse. Beneath this abstract I-O relation is a more detailed description of the intermediate stages in this mechanism: the tetanus excites the postsynaptic cell (A) which depolarizes it (B), causes NMDA receptors to open (C), and so on. (Nothing turns on the fact that I’ve idealized this mechanism as a single, step-wise causal chain.) Now, suppose we intervene to open the NMDA receptors directly and thereby potentiate the synapse (as shown in the third line). We might say that this intervention induced LTP; but when we say this, we mean that it produces the end product of the mechanism (it potentiates the synapse). We cannot coherently assert that it causes the process as depicted in the first two lines. This is for the simple reason that the process in the first two lines includes stages I, A, and B, and these are absent in the causal sequence represented in the third line. At most we can say the intervention caused the last half of the process. The NMDA receptor is not a cause of the process of LTP; it is a component of that process.[27]