David Danks

Bayesian models of human learning are becoming increasingly popular in cognitive science. We argue that their purported confirmation largely relies on a methodology that depends on premises that are inconsistent with the claim that people are Bayesian about learning and inference. Bayesian models in cognitive science derive their appeal from their normative claim that the modeled inference is in some sense rational. Standard accounts of the rationality of Bayesian inference imply predictions that an agent selects the option that maximizes the posterior expected utility. Experimental confirmation of the models, however, has been claimed because of groups of agents that probability match the posterior. Probability matching only constitutes support for the Bayesian claim if additional unobvious and untested (but testable) assumptions are invoked. The alternative strategy of weakening the underlying notion of rationality no longer distinguishes the Bayesian model uniquely. A new account of rationality—either for inference or for decision-making—is required to successfully confirm Bayesian models in cognitive science.

In many people, caffeine causes slight muscle tremors, particularly in their hands. In general, the Caffeine → Muscle Tremors causal connection is a noisy one: someone can drink coffee and experience no hand shaking, and there are many other factors that can lead to muscle tremors. Now suppose that Jane drinks several cups of coffee and then notices that her hands are trembling; an obvious question is: did this instance of coffee drinking cause this instance of hand-trembling? Structurally similar questions arise throughout everyday life: Did this pressing of the ‘k’ key cause this change in the pixels on the computer monitor? Did these episodes of smoking cause this lung cancer? Did this studying cause this test score? And so on. These questions all ask about singular causation in a particular situation, in contrast with general causation across multiple cases. They are thus particularly salient in situations in which we care about that specific case, as in many legal contexts, social interactions, physical explanations of anomalous events, and more

Our concept of actual causation plays a deep, ever-present role in our experiences. I first argue that traditional philosophical methods for understanding this concept are unlikely to be successful. I contend that we should instead use functional analyses and an understanding of the cognitive bases of causal cognition to gain insight into the concept of actual causation. I additionally provide initial, programmatic steps towards carrying out such analyses. The characterization of the concept of actual causation that results is quite different from many standard views: it is graded, context-sensitive, and extrinsic

There is now substantial agreement about the representational component of a normative theory of causal reasoning: Causal Bayes Nets. There is less agreement about a normative theory of causal discovery from data, either computationally or cognitively, and almost no work investigating how teaching the Causal Bayes Nets representational apparatus might help individuals faced with a causal learning task. Psychologists working to describe how naïve participants represent and learn causal structure from data have focused primarily on learning from single trials under a variety of conditions. In contrast, one component of the normative theory focuses on learning from a sample drawn from a population under some experimental or observational study regime. Through a virtual Causality Lab that embodies the normative theory of causal reasoning and which allows us to record student behavior, we have begun to systematically explore how best to teach the normative theory. In this paper we explain the overall project and report on pilot studies which suggest that students can quickly be taught to (appear to) be quite rational

Erratum to: Synthese DOI 10.1007/s11229-014-0408-3Appendix 1: NotationLet \(X\) represent a sequence of data, and let \(X_B^t\) represent an i.i.d. subsequence of length \(t\) of data generated from distribution \(B\).We conjecture that the i.i.d. assumption could be eliminated by defining probability distributions over sequences of arbitrary length, though this complication would not add conceptual clarity. Let \(\mathbf{F}\) be a framework (in this case, a set of probability distributions or densities).Let any \(P(\,)\) functions be either a probability distribution function or probability density function, as appropriate. Let \(M_\mathbf{F}\) be a method that takes a data sequence \(X\) as input and outputs a distribution

Even if one can experiment on relevant factors, learning the causal structure of a dynamical system can be quite difficult if the relevant measurement processes occur at a much slower sampling rate than the “true” underlying dynamics. This problem is exacerbated if the degree of mismatch is unknown. This paper gives a formal characterization of this learning problem, and then provides two sets of results. First, we prove a set of theorems characterizing how causal structures change under undersampling. Second, we develop an algorithm for inferring aspects of the causal structure at the “true” timescale from the causal structure learned from the undersampled data. Research on causal learning in dynamical contexts has largely ignored the challenges of undersampling, but this paper provides a framework and foundation for learning causal structure from this type of complex time series data

We evaluate the asymptotic performance of boundedly-rational strategies in multi-armed bandit problems, where performance is measured in terms of the tendency (in the limit) to play optimal actions in either (i) isolation or (ii) networks of other learners. We show that, for many strategies commonly employed in economics, psychology, and machine learning, performance in isolation and performance in networks are essentially unrelated. Our results suggest that the appropriateness of various, common boundedly-rational strategies depends crucially upon the social context (if any) in which such strategies are to be employed.

We argue that current discussions of criteria for actual causation are ill-posed in several respects. (1) The methodology of current discussions is by induction from intuitions about an infinitesimal fraction of the possible examples and counterexamples; (2) cases with larger numbers of causes generate novel puzzles; (3) "neuron" and causal Bayes net diagrams are, as deployed in discussions of actual causation, almost always ambiguous; (4) actual causation is (intuitively) relative to an initial system state since state changes are relevant, but most current accounts ignore state changes through time; (5) more generally, there is no reason to think that philosophical judgements about these sorts of cases are normative; but (6) there is a dearth of relevant psychological research that bears on whether various philosophical accounts are descriptive. Our skepticism is not directed towards the possibility of a correct account of actual causation; rather, we argue that standard methods will not lead to such an account. A different approach is required

A pervasive feature of the sciences, particularly the applied sciences, is an experimental focus on a few (often only one) possible causal connections. At the same time, scientists often advance and apply relatively broad models that incorporate many different causal mechanisms. We are naturally led to ask whether there are normative rules for integrating multiple local experimental conclusions into models covering many additional variables. In this paper, we provide a positive answer to this question by developing several inference rules that use local causal models to place constraints on the integrated model, given quite general assumptions. We also demonstrate the practical value of these rules by applying them to a case study from ecology. Experimental scope in applied sciences Fusing the results of experiments A concrete example of the inference rules Application to a case study.

Arguments, claims, and discussions about the “level of description” of a theory are ubiquitous in cognitive science. Such talk is typically expressed more precisely in terms of the granularity of the theory, or in terms of Marr’s three levels. I argue that these ways of understanding levels of description are insufficient to capture the range of different types of theoretical commitments that one can have in cognitive science. When we understand these commitments as points in a multi-dimensional space, we find that we must also reconsider our understanding of intertheoretic relations. In particular, we should understand cognitive theories as constraining one another, rather than reducing to one another

Our best sciences are frequently held to be one way, perhaps the optimal way, to learn about the world’s higher-level ontology and structure. I first argue that which scientific theory is “best” depends in part on our goals or purposes. As a result, it is theoretically possible to have two scientific theories of the same domain, where each theory is best for some goal, but where the two theories posit incompatible ontologies. That is, it is possible for us to have goal-dependent pluralism in our scientific ontologies. This ontological pluralism arises simply from our inability to directly know the world’s objects, rather than any particular claims about our cognitive limits, values, or social structures. I then present two case studies in which this possibility actually occurs—one based on simulations and theoretical analyses of constructed causal systems, and one from actual scientific investigations into the proper ontology for ocean regions

Research on adaptive rationality has focused principally on inference, judgment, and decision-making that lead to behaviors and actions. These processes typically require cognitive representations as input, and these representations must presumably be acquired via learning. Nonetheless, there has been little work on the nature of, and justification for, adaptively rational learning processes. In this paper, we argue that there are strong reasons to believe that some learning is adaptively rational in the same way as judgment and decision-making. Indeed, overall adaptive rationality can only properly be assessed for pairs of learning and decision processes. We thus present a formal framework for modeling such pairs of cognitive processes, and thereby assessing their adaptive rationality relative to the environment and the agent’s goals. We then use this high-level formal framework on specific cases by demonstrating how natural formal constraints on decision-making can lead to substantive predictions about adaptively rational learning and representation; and characterizing adaptively rational learning for fast-and-frugal one-reason decision-making.

One persistent challenge in scientific practice is that the structure of the world can be unstable: changes in the broader context can alter which model of a phenomenon is preferred, all without any overt signal. Scientific discovery becomes much harder when we have a moving target, and the resulting incorrect understandings of relationships in the world can have significant real-world and practical consequences. In this paper, we argue that it is common (in certain sciences) to have changes of context that lead to changes in the relationships under study, but that standard normative accounts of scientific inquiry have assumed away this problem. At the same time, we show that inference and discovery methods can “protect” themselves in various ways against this possibility by using methods with the novel methodological virtue of “diligence.” Unfortunately, this desirable virtue provably is incompatible with other desirable methodological virtues that are central to reliable inquiry. No scientific method can provide every virtue that we might want

Newsome ((2003). The debate between current versions of covariation and mechanism approaches to causal inference. Philosophical Psychology, 16, 87-107.) recently published a critical review of psychological theories of human causal inference. In that review, he characterized covariation and mechanism theories, the two dominant theory types, as competing, and offered possible ways to integrate them. I argue that Newsome has misunderstood the theoretical landscape, and that covariation and mechanism theories do not directly conflict. Rather, they rely on distinct sets of reliable indicators of causation, and focus on different types of causation (type vs. token). There are certainly debates in the research field, but the theoretical landscape is not as fractured as Newsome suggests, and a potential unifying framework has already emerged using causal Bayes nets. Philosophical work on causal epistemology matters for psychologists, but not in the way Newsome suggests.

Standard time series structure learning algorithms assume that the measurement timescale is approximately the same as the timescale of the underlying system. In many scientific contexts, however, this assumption is violated: the measurement timescale can be substantially slower than the system timescale. This assumption violation can lead to significant learning errors. In this paper, we provide a novel learning algorithm to extract systemtimescale structure from measurement data that undersample the underlying system. We employ multiple algorithmic optimizations that exploit the problem structure in order to achieve computational tractability. The resulting algorithm is highly reliable at extracting system-timescale structure from undersampled data

Debates about the moral permissibility of mass electronic surveillance often turn on whether consequentialist considerations legitimately trump relevant deontological rights and principles. In order to establish such overriding consequences, many proponents of mass surveillance employ a modern analogue of Pascal’s wager: they contend that the consequences of no surveillance are so severe that any probability of such outcomes legitimates the abrogation of the relevant rights. In this paper, I briefly review Pascal’s original wager about whether to live a pious life, including two classes of objections that were almost immediately leveled against his argument. I then show that analogues of those objections apply straightforwardly to the modern versions of Pascal’s wager. Mass electronic surveillance might be ethically permissible or even obligatory in some circumstances, but the details matter in ways that are systematically ignored or assumed away in a Pascal’s wager-type argument. Careful consideration of the complexities of our modern situation is required to decide whether consequentialist considerations override deontological principles in this domain

In everyday matters, as well as in law, we allow that someone’s reasons can be causes of her actions, and often are. That correct reasoning accords with Bayesian principles is now so widely held in philosophy, psychology, computer science and elsewhere that the contrary is beginning to seem obtuse, or at best quaint. And that rational agents should learn about the world from energies striking sensory inputs nerves in people—seems beyond question. Even rats seem to recognize the difference between correlation and causation,1 and accordingly make different inferences from passive observation than from interventions. A few statisticians aside,” so do most of us. To square these views with the demands of computability, increasing numbers of psychologists and others have embraced a particular formalization, causal Bayes nets, as an account of human reasoning about and to causal connections.111 Such structures can be used by rational agents, including humans in so far as they are rational, to have degrees of belief in various conceptual contents, which they use to reason to expectations, which are realized or defeated by sensory inputs, which cause them to change their degrees of belief in other contents in accord with Bayes Rule, or some generalization of it. How is all of this supposed to be carried out? l. Representing Causal Structures The causal Bayes net framework adopted by a growing number of psychologists goes like this: Our representations of causal relations are captured in a graphical causal.

Models based on causal capacities, or independent causal influences/mechanisms, are widespread in the sciences. This paper develops a natural mathematical framework for representing such capacities by extending and generalizing previous results in cognitive psychology and machine learning, based on observations and arguments from prior philosophical debates. In addition to its substantial generality, the resulting framework provides a theoretical unification of the widely-used noisy-OR/AND and linear models, thereby showing how they are complementary rather than competing. This unification helps to explain many of the shared cognitive and mathematical properties of those models.