It is often claimed that one cannot locate a notion of causation in fundamental physical theories. The reason most commonly given is that the dynamics of those theories do not support any distinction between the past and the future, and this vitiates any attempt to locate a notion of causal asymmetry—and thus of causation—in fundamental physical theories. I argue that this is incorrect: the ubiquitous generation of entanglement between quantum systems grounds a relevant asymmetry in the dynamical evolution of quantum (...) systems. I show that by exploiting a connection between the amount of entanglement in a quantum state and the algorithmic complexity of that state, one can use recently developed tools for causal inference to identify a causal asymmetry—and a notion of causation—in the dynamical evolution of quantum systems. (shrink)

Causal representations are distinguished from non-causal ones by their ability to predict the results of interventions. This widely-accepted view suggests the following adequacy condition for causal models: a causal model is adequate only if it does not contain variables regarding which it makes systematically false predictions about the results of interventions. Here I argue that this condition should be rejected. For a class of equilibrium systems, there will be two incompatible causal models depending on whether one intervenes upon a certain (...) variable to fix its value, or ‘lets go’ of the variable and allows it to vary. The latter model will fail to predict the result of interventions on the let-go-of variable. I argue that there is no basis for preferring one of these models to the other, and thus that models failing to predict interventions on particular variables can be just as adequate as those making no such false predictions. This undermines a key argument against relying upon causal models inferred from equilibrium data. (shrink)

Causal representations are distinguished from non-causal ones by their ability to predict the results of interventions. This widely-accepted view suggests the following adequacy condition for causal models: a causal model is adequate only if it does not contain variables regarding which it makes systematically false predictions about the results of interventions. Here I argue that this condition should be rejected. For a class of equilibrium systems, there will be two incompatible causal models depending on whether one intervenes upon a certain (...) variable to fix its value, or ‘lets go’ of the variable and allows it to vary. The latter model will fail to predict the result of interventions on the let-go-of variable. I argue that there is no basis for preferring one of these models to the other, and thus that models failing to predict interventions on particular variables can be just as adequate as those making no such false predictions. This undermines a key argument (Dash in Caveats for causal reasoning with equilibrium models. University of Pittsburgh. PhD thesis, 2003) against relying upon causal models inferred from equilibrium data. (shrink)

We propose a formal framework for interpretable machine learning. Combining elements from statistical learning, causal interventionism, and decision theory, we design an idealised explanation game in which players collaborate to find the best explanation for a given algorithmic prediction. Through an iterative procedure of questions and answers, the players establish a three-dimensional Pareto frontier that describes the optimal trade-offs between explanatory accuracy, simplicity, and relevance. Multiple rounds are played at different levels of abstraction, allowing the players to explore overlapping causal (...) patterns of variable granularity and scope. We characterise the conditions under which such a game is almost surely guaranteed to converge on a optimal explanation surface in polynomial time, and highlight obstacles that will tend to prevent the players from advancing beyond certain explanatory thresholds. The game serves a descriptive and a normative function, establishing a conceptual space in which to analyse and compare existing proposals, as well as design new and improved solutions. (shrink)

We propose a formal framework for interpretable machine learning. Combining elements from statistical learning, causal interventionism, and decision theory, we design an idealisedexplanation gamein which players collaborate to find the best explanation(s) for a given algorithmic prediction. Through an iterative procedure of questions and answers, the players establish a three-dimensional Pareto frontier that describes the optimal trade-offs between explanatory accuracy, simplicity, and relevance. Multiple rounds are played at different levels of abstraction, allowing the players to explore overlapping causal patterns of (...) variable granularity and scope. We characterise the conditions under which such a game is almost surely guaranteed to converge on a (conditionally) optimal explanation surface in polynomial time, and highlight obstacles that will tend to prevent the players from advancing beyond certain explanatory thresholds. The game serves a descriptive and a normative function, establishing a conceptual space in which to analyse and compare existing proposals, as well as design new and improved solutions. (shrink)

In the Origin, Darwin forwards two incompatible lines of attack on special creationism. First, he argues that imperfect or functionless traits are evidence against design. Second, he argues that since special creationism can be made compatible with any observation, it is unscientific and explanatorily vacuous. In later works, Darwin shifts to an argument that he finds much more persuasive and which would undermine theistic evolutionism as well. He argues that variation is random with respect to selection and that this demonstrates (...) that there is no design in the biological world. I examine why Darwin found the argument from independence of variation and selection more compelling than the argument from imperfection. I argue that Darwin utilized general principles of causal inference, akin to those used in modern causal modeling, to rule out any unified cause behind the evolutionary process. (shrink)

Functional robustness refers to a system’s ability to maintain a function in the face of perturbations to the causal structures that support performance of that function. Modularity, a crucial element of standard methods of causal inference and difference-making accounts of causation, refers to the independent manipulability of causal relationships within a system. Functional robustness appears to be at odds with modularity. If a function is maintained despite manipulation of some causal structure that supports that function, then the relationship between that (...) structure and function fails to be manipulable independent of other causal relationships within the system. Contrary to this line of reasoning, I argue that functional robustness often attends feedback control, rather than failures of modularity. Feedback control poses its own challenges to causal explanation and inference, but those challenges do not undermine modularity—and indeed, modularity is crucial to grappling with them. (shrink)

The reduction of grounding to causation, or each to a more general relation of which they are species, has sometimes been justified by the impressive inferential capacity of structural equation modelling, causal Bayes nets, and interventionist causal modelling. Many criticisms of this assimilation focus on how causation is inadequate for grounding. Here, I examine the other direction: how treating grounding in the image of causation makes the resulting view worse for causation. The distinctive features of causal modelling that make this (...) connection appealing are distorted beyond use by forcing them to fit onto grounding. The very inferential strength that makes causation attractive is only possible because of a narrow construal of what counts as a causal relation; as soon as that broadens, the inferential capacity markedly diminishes. Making causation suitable for application to grounding spoils what was appealing about causation for this task in the first place. However, grounding need not appeal to causation: causal modelling does not have exclusive claim to structural equation modeling or other formal techniques of modelling structure. I offer a case in favour of a different kind of metaphysical frugality, which tend towards narrow, more restrictive construals of relations like causation or grounding, because then each relation behaves more homogenously. This more homogenous behavior delivers stronger inferential power per relation even though there may be more relations to which one is committed. (shrink)

I highlight a metaphysical concern that stands in the way of more widespread adoption of causal modeling techniques such as causal Bayes nets. Researchers in some fields may resist adoption due to concerns that they don't 'really' understand what they are saying about a system when they apply such techniques. Students in these fields are repeated exhorted to be cautious about application of statistical techniques to their data without a clear understanding of the conditions required for those techniques to yield (...) genuine insight into the data. They are acutely aware that anyone can chuck some data into a software package and get what looks like an answer, even though these tests may not be well-defined for the data on which they can apparently be run. This is thus a healthy skepticism for uptake of causal modeling methods, which points directly to the need for a metaphysical understanding of causation in order to successfully use the modeling methods. Without a clear understanding of what the methods are committing to, including what it means to say that there exists a causal relationship, researchers have limited ability to identify potentially bad output, or to independently verify results. (shrink)