Abstracts

Mark Couch (Seton Hall University) – ‘How to model determination relations using INUS conditions’

It is commonly said that mechanisms are constitutively related to the phenomena they explain. While this view is held by many (Craver 2007, Baumgartner and Casini 2017), the nature of this constitutive relation has remained obscure, with some even claiming that it is better characterized as a causal relation. In this talk I will defend the view that the relation between mechanisms and their phenomena should be characterized constitutively. This approach to mechanisms fits with the fact that the relation between mechanisms and phenomena is synchronic. Moreover, the best developed account of the relation is described by Craver and Wilson (2007), who characterize constitution in terms of a relation of metaphysical sufficiency. This notion is explicitly noncausal and is intended to be similar to relations like “realization” in the literature. In this respect the two relations form a group of distinct, noncausal determination relations.

After this I will then develop an account of a constitutively relevant part of a mechanism. This account draws on previous work from Couch (2011) and Harbecke (2010), who’s theory of mechanistic constitution characterizes relevant parts in terms of inus conditions. The idea is that a part of a mechanism is constitutively relevant to a phenomenon of a mechanism just in case the part is inus for the phenomenon. I will present this characterization of the relevant parts of mechanisms and explain how it fits into the account of constitution as a sufficiency relation described above.

In the remainder of the talk I will go on to develop this account more broadly in ways that have not yet been explored. The aim will be to show how the inus notion can be used to characterize a variety of determination relations and is broader than usually understood. One of the interesting features of Mackie’s (1965) notion of inus conditions, note, is that he claims it is distinct from the particular analysis of necessity he offers in his work. His point is that this notion characterizes a form of causal determination that can be incorporated into various ways of thinking about causal relations. I will develop this point to show how this works for a variety of cases. In particular, I will show that we can characterize two types of determination relations using an inus notion: causal determination and constitutive determination. The trick is to see that in each case the relevant notion of “necessary and sufficient” conditions is characterized along different modal dimensions. This modal characterization goes beyond the regularity characterizations of inus conditions that have been offered elsewhere.

I will develop this approach and explain how it relates to the nature of mechanisms. In doing this I will also consider work from Strevens (2007) that claims Mackie’s notion can be refined and used in the context of theories that Mackie wouldn’t accept. This is important to note since it helps explain the flexibility of Mackie’s original notion. Exploring the different ways of understanding such determination claims will, thus, help us understand these claims more generally, and in a way different from the interventionist account.

References:

Baumgartner M. and Casini L. (2017) “An Abductive Theory of Constitution,” Philosophy of Science, 84, 214-233.
Couch, M. (2011) “Mechanisms and Constitutive Relevance,” Synthese, 183, 375-388.
Craver C. (2007) Explaining the Brain, NY: Oxford University Press. 
Craver C. and Wilson, R. (2007) “Realization,” In P. Thagard (ed), Handbook of the Philosophy of Psychology and Cognitive Science, 81-104, Kluwer.
Harbecke, J. (2010) “Mechanistic Constitution in Neurobiological Explanations,” International Studies in the Philosophy of Science, 24, 267-285.
Mackie, J. L. (1965) “Causes and Conditions,” in E. Sosa & M. Tooley (eds) (1993), Causation, 33-55, NY: Oxford University Press.
Strevens, M. (2007) “Mackie Remixed,” in J. Campbell, M. O’Rourke, & M. Silverstein, Causation and Explanation, 93-118, Cambridge, MA: MIT Press.
 

Frederick Eberhardt (California Institute of Technology) – ‘Causal macro variables’

Standard methods of causal discovery take as input a statistical data set of measurements of well-defined causal variables. The goal is then todetermine the causal relations among these variables. But how are these causal variables identified or constructed in the first place? Often we have sensor level data but assume that the relevant causal interactions occur at a higher scale of aggregation. Sometimes we only have aggregate measurements of causal interactions at a finer scale. I will motivate the general problem of causal discovery and present recent work on a framework and method for the construction and identification of causal macro-variables that ensures that the resulting causal variables have well-defined intervention distributions.

Markus Eronen (University of Groningen) - 'Discovering constitutive relevance through interventions: Obstacles and challenges'

Many recently proposed definitions of constitutive relevance rely oninterventions. In this talk, I argue that in psychology (and neuroscience)interventions are typically soft and fat-handed, and thus very far removedfrom the Woodwardian regulative ideal. In addition, due to the nature of psychological measurement, the degree to which a psychological interventio nwas soft and fat-handed, or more generally, what the intervention in fact did, is difficult to reliably estimate. These problems create a formidable obstacle to intervention-based discovery of constitutive relationships in psychology (and neuroscience). I also consider to what extent they indirectly apply to non-interventionist accounts of constitutive relevance.

Peter Fazekas (University of Antwerp and Aarhus University) – ‘Fat hands are not that fat in a flat world - Constitution and mutual manipulability in flat mechanisms’

Theories of mechanistic explanation need to be able to determine which entities are the working parts — i.e. constitutively relevant for maintaining the functions — of any given mechanism. Traditionally, the criterion used for this purpose is mutual manipulability. Roughly, an entity is a working part of a mechanism if there is an ideal intervention on the entity under which the overall behaviour of the whole mechanism changes, and there is an ideal intervention on the whole under which the activity of the entity changes.

In recent literature, this criterion has been criticised on the basis that ideal interventions are impossible at higher levels without also intervening at lower levels. An ideal intervention on a whole with respect to a component entity would induce a change in the whole without having an immediate impact on the component. However, since whole mechanisms supervene on their parts and the parts’ activities, and there cannot be a change in the supervening property without there being a change in at least one of the subvening properties, every intervention on the whole is necessarily ‘fat handed’, i.e. it will necessarily induce a change in at least one of the parts as well. So within the mechanistic framework, ideal interventions on wholes with respect to their parts are not possible, and hence the mutual manipulability criterion cannot be satisfied. The reigning conclusion drawn from this line of thought is that non-reductive physicalism and interventionism together are incompatible with the idea of mutual manipulability. In this paper, I demonstrate that it is, in fact, not the case — that what causes the problem is the levelled view of reality that the mechanistic framework traditionally comes bundled with, according to which different entities forming part-whole relations reside at lower and higher levels, respectively. I develop an alternative flat view, and show that within this flat view non-reductive physicalism and interventionism are compatible with mutual manipulability.

First, to motivate the flat view, I argue that contrary to the claims of theproponents of the mechanistic framework, its core commitments are in fact incompatible with the levelled view. Then I develop the flat view, according to which wholes do not belong to levels higher than the constituent parts of the underlying mechanisms, but rather are to be found as modules embedded in the very same complex of interacting units. Modules are structurally and functionally stable configurations of the interacting units composing them, and are encapsulated either in a direct physical way by a boundary that separates them from their environment, or functionally by the specific organisation of the interaction network of their units (e.g. causal feedback loops). Physical and functional encapsulation constrain internal operations, cut off some internal-external interactions, and screen off inner organisation and activities. Due to the cutting-off effect of encapsulation, the interacting units of a module are, to a certain degree, causally detached from their environment: some of the causal paths via which the units could normally (in separation) be influenced become either unavailable (due to the shielding effect of physical boundaries) or ineffective (due to the stabilising effect of feedback loops). Some units, however, still retain their causal links with the environment providing inputs (input units) and outputs (output units) for the organised activity of the cluster of units, and henceforth for the module itself. Via the causal links of their input and output units, modules are causallyembedded in the same level of causal interactions as their component units. Sincewhole modules can be influenced by and can influence their environment only via their input and output units, their inner organisation is screened off: from the ‘outside’ modules function as individual units. Therefore, alternating between a module and a unit view is only a change in perspective: the mechanistic programme consists in turning units into modules, i.e. ‘blowing up’ the unit under scrutiny to uncover its internal structure, and accounting for its behaviour in terms of the organisation and activities of the units found ‘inside’. The flat view, thus, claims that mechanistic characterisa- tions of different ‘levels’ are to be understood as different descriptions providing different levels of detail with regard to a set of interacting units with complex embedded structure. Moreover, the flat view is compatible with non-reductive physicalism, since it can accommodate multiple realisability and epistemological emergence as the same input-output functions can be maintained by different internal organisation.

Within the flat view, demarcating a mechanism amounts to determining which units belong to a module. Since modules are functionally identified via their input-output (cause-effect) functions, the task is to determine which units play a part in processing the input of the module and transforming it into its output. Wholes are connected to their environment with their input and output channels, so intervening on a whole amounts to affecting its input channels, i.e. interacting with its input units, whereas detecting how the behaviour of the module changes amounts to detecting the output of the module's output units. Mutual manipulability within the flat view, thus, works as follows. Figure 1 shows an exemplar module with a couple of units. IN is the input unit, OUT is the output unit. The question is whether unit X belongs to the module. This can be determined by intervening on the whole, i.e. intervening on IN, and then looking for a change in X; and intervening on X and then looking for a change in the whole, i.e. looking for a change in OUT. Neither the change in IN and the change in X (Fig. 1a), nor the change in X and the change in OUT (Fig. 1b) are in non-causal determination relation with each other — the only link connecting them is purely causal. So within the flat view, interventionism is on safe grounds, and ideal interventions are possible. Therefore, mutual manipulability is unproblematic within the flat view: it can serve its original purpose and help find constitutively relevant parts and demarc ate mechanisms from their environments.

[figures shown in the sidebar]

Alexander Gebharter (University of Groningen) and Jens Harbecke (Witten/Herdecke University) – ‘Constitutive relevance discovery without interventions: Boole meets Bayes’

In recent years, the topic of inferring relationships of constitutive relevance has received considerable attention within the debate on mechanistic explanation. Most accounts currently on the market aim at identifying such relationships on the basis of systematic interventions. In this paper, in contrast, we focus on methods for constitutive discovery that do not rely on interventions. In particular, we explore the specific strengths and weaknesses of the Boolean (Harbecke, 2015) and the Bayesian (Gebharter, 2017) approach and argue that certain strategies are available to combine and synthesize the two methods formally and heuristically.

David Kinney (LSE) – ‘Pragmatic causal feature learning’

I propose a pragmatic model of causal feature learning. This model defines a procedure for building coarse-grained macro-variables from more fine-grained micro-variables such that the macro-variables supervene on the micro-variables, and such that causal relationships between variables are preserved during the coarse-graining procedure. The approach is pragmatic in the sense that it does not permit the loss of valuable information, where the value of information is defined in a decision-theoretic way that is indexed to a particular decision problem faced by a particular agent. This stands in contrast to purely epistemic models of causal feature learning developed by Chalpuka, Eberhardt, and Perona. I also draw some connections to the work of Casini et al.

Beate Krickel (Ruhr University Bochum) – ‘Mechanistic constitution - How many relations?’

Plausibly, the engine's running is a component of the mechanism that explains the car's moving. Also, the wheel's turning is plausibly a component in that mechanism. In the terminology of the new mechanists, thus, both at least partially constitute the car's moving. Despite their plausibility, these intuitive cases create a tension: the first example suggests that constitution is an asynchronous relation as any relevant change in the engine’s running will affect the car’s moving only at a later time. In contrast to that, the second example suggest that constitution is a synchronous relation as any relevant change in the wheel’s turning will simultaneously change the car’s moving.

In this talk, I present two accounts of constitutive relevance that either depict constitution as involving a temporal delay (Krickel 2018), or as involving simultaneous changes (Baumgartner Casini 2017; Baumgartner, Casini, Krickel 2018). I discuss whether the two accounts can be integrated or whether we are indeed dealing with two different relations rather than one.

Bert Leuridan (University of Antwerp) – ‘Constitution versus causation - What after the redux?’

In the past ten years, Carl Craver's account of complex-system mechanisms -- in particular his views on levels, constitutive relations and interlevel causation -- has attracted widespread philosophical interest. Much of this interest has centered on (i) his mutual manipulability account of constitutive relevance and (ii) what have come to be known as 'Craver-diagrams', pie-tin diagrammatic representations of multi-level mechanisms. Recently, however, Craver has (i*) offered a new account of mutual manipulability and (ii*) abandoned the use of pie-tin diagrams in favour of flat input-output graphs in a paper entitled "Mutual Manipulability Redux". This paper may well be a game changer in the literature. I will analyze the consequences which it may have for existing arguments regarding the relations between (a) interventionism, (b) constitutive relevance and (c) interlevel causation. Note: as Craver's Redux paper is still under construction and under embargo, my conclusions will be very tentative and exploratory.

Daniel Malinsky (Johns Hopkins University) – 'Learning about changes to causal structure'

Functional causal models or structural equation models (SEMs) relate random variables to their causes by deterministic functions (and stochastic“noise” terms). Once we’ve come to believe some such functional relationship holds for particular variables Y, X, … our interests might turn to answering counterfactual questions (or w-questions) involving hypothetical changes to those functional relationships. What would be the distribution of Y if the causal dependence of Y on X were weaker, perhaps zero? In Malinsky (2018) I argued both that 1) we can fruitfully think of such questions as queries about a post-intervention distribution, where the intervention modifies the functional form or structural parameters, and 2)that by doing so we can explicate debates over the causal or explanatory role of “macro” structural features of a system under study, which seem notto correspond to a single variable in any realistic model. For example,discussions of systemic racism, macroeconomic organization (e.g., “command economy” versus “free market”), or heterogeneity of biological mechanisms may be understood on such terms. There is an attendant epistemological question: how could we actually learn about such interventions on causal structure? In this talk, I’ll dive more deeply into the epistemology. I’ll rephrase the original proposal non-parametrically, and consider how statistical techniques recently deployed in a very different setting —algorithmic fairness — can perhaps be leveraged to make some w-questions about causal structure estimable from statistical data.

Ref: Malinsky, D. (2018) “Intervening on structure.” Synthese 195(5),2295-2312.

Alessio Moneta (Scuola Superiore Sant’anna) - 'Independent components and the causation-constitution distinction'

Theories of mechanistic explanation hold that an explanandum phenomenon non-reductively supervenes on, and is explained by, its constituting entities and activities. Since constitution is the core dependence relation in such explanations, theories of mechanistic explanation require a method for discovering constitutional relations. In particular, mechanistic explanations are “downward-looking”: they explain upper-level phenomena in terms of their lower-level constituents. They differ from causal explanations, which explain effects in terms of their causes, and are thus “backward-looking”. To make sense of this distinction, a suitable method for constitutional discovery should allow one to distinguish constitutive from causal dependencies. Here, we argue that recovering an independent component (IC) representation of the data suffices to identify constitutive dependencies and distinguish them from causal dependencies, provided the variable set includes all constituents of a given phenomenon, or it contains a subset of (observed) constituents as well as one effect for each of the other (unobserved) constituents. Under these conditions, our proposal rationalizes the difference between causal and constitutive (or mechanistic) explanations.

Maria Serban (Technische Universität Berlin) – ‘Reconstruction (not interlevel) experiments‘

One important aim of experimentation in molecular and cellular cognition is to identify the processes via which memories are formed and stored for short or long term use. A number of mechanistic philosophers have argued that learning or "memory consolidation" experiments together with synaptic plasticity experiments reveal the multilvel structure of these mechanisms (Craver and Darden 2001; Craver 2002, 2007). And vice versa, that the multilevel organization of mechanisms shapes the experimental procedures by which they are discovered. Thus experimental methodology in fields like cognitive neurobiology is said to be committed to both causal and constitutive discovery/inference. However, one of the most popular interpretations of constitutive discovery, the mutual manipulability (MM)account (Craver 2007), has come under severe scrutiny (Baumgartner and Gebharter 2016; Baumgartner and Casini 2017). The culprit of many problems raised for the MM account is the structure of so-called "top-down" experiments. I argue that some of these experiments are best understood as "reconstruction" experiments, aiming to identify and establish the construct validity of a specific form of learning or synaptic plasticity, while others are experiments showing that some target factor or variable is a causal intermediary in a specific mechanism (Baetu 2012). This interpretation of the structure of experimentation in molecular and cellular cognition is partly in line with the interventionist intuitions of the MM account but it does not postulate any additional commitments to interlevel mechanistic dependence relations.

Jiji Zhang (Lingnan University) – ‘Structural equation models for both causation and constitution’

In this talk I generalize Judea Pearl's framework of modifiable structural equation models to accommodate non-causal as well as causal dependencies. I illustrate the utility of the generalized framework by relating it to a host of important issues in philosophy and in machine learning, including interventionist conditionals, supervenient causation, ambiguous intervention effects, and macro variable construction.