Current psychological theories of human causal learning and judgment focus primarily on long-run predictions: two by estimating parameters of a causal Bayes nets, and a third through structural learning. This paper focuses on people’s short-run behavior by examining dynamical versions of these three theories, and comparing their predictions to a real-world dataset.

The goal of the present set of studies is to explore the boundary conditions of category transfer in causal learning. Previous research has shown that people are capable of inducing categories based on causal learning input, and they often transfer these categories to new causal learning tasks. However, occasionally learners abandon the learned categories and induce new ones. Whereas previously it has been argued that transfer is only observed with essentialist categories in which the (...) hidden properties are causally relevant for the target effect in the transfer relation, we here propose an alternative explanation, the unbroken mechanism hypothesis. This hypothesis claims that categories are transferred from a previously learned causal relation to a new causal relation when learners assume a causal mechanism linking the two relations that is continuous and unbroken. The findings of two causal learning experiments support the unbroken mechanism hypothesis. (shrink)

Causal understanding is a defining characteristic of human cognition. Like many animals, human children learn to control their bodily movements and act effectively in the environment. Like a smaller subset of animals, children intervene: they learn to change the environment in targeted ways. Unlike other animals, children grow into adults with the causal reasoning skills to develop abstract theories, invent sophisticated technologies and imagine alternate pasts, distant futures and fictional worlds. In this Review, we explore the development of (...) human-unique causal learning and reasoning from evolutionary and ontogenetic perspectives. We frame our discussion using an ‘interventionist’ approach. First, we situate causal understanding in relation to cognitive abilities shared with non-human animals. We argue that human causal understanding is distinguished by its depersonalized (objective) and decontextualized (general) representations. Using this framework, we next review empirical findings on early human causal learning and reasoning and consider the naturalistic contexts that support its development. Then we explore connections to related abilities. We conclude with suggestions for ongoing collaboration between developmental, cross-cultural, computational, neural and evolutionary approaches to causal understanding. (shrink)

We propose that children employ specialized cognitive systems that allow them to recover an accurate “causal map” of the world: an abstract, coherent, learned representation of the causal relations among events. This kind of knowledge can be perspicuously understood in terms of the formalism of directed graphical causal models, or “Bayes nets”. Children’s causal learning and inference may involve computations similar to those for learning causal Bayes nets and for predicting with them. Experimental (...) results suggest that 2- to 4-year-old children construct new causal maps and that their learning is consistent with the Bayes net formalism. (shrink)

Occam’s razor directs us to adopt the simplest hypothesis consistent with the evidence. Learning theory provides a precise definition of the inductive simplicity of a hypothesis for a given learning problem. This definition specifies a learning method that implements an inductive version of Occam’s razor. As a case study, we apply Occam’s inductive razor to causal learning. We consider two causal learning problems: learning a causal graph structure that presents global (...) class='Hi'>causal connections among a set of domain variables, and learning context-sensitive causal relationships that hold not globally, but only relative to a context. For causal graph learning, Occam’s inductive razor directs us to adopt the model that explains the observed correlations with a minimum number of direct causal connections. For expanding a causal graph structure to include context-sensitive relationships, Occam’s inductive razor directs us to adopt the expansion that explains the observed correlations with a minimum number of free parameters. This is equivalent to explaining the correlations with a minimum number of probabilistic logical rules. The paper provides a gentle introduction to the learning-theoretic definition of inductive simplicity and the application of Occam’s razor for causal learning. (shrink)

Two key research issues in the field of causal learning are how people acquire causal knowledge when observing data that are presented sequentially, and the level of abstraction at which learning takes place. Does sequential causal learning solely involve the acquisition of specific cause-effect links, or do learners also acquire knowledge about abstract causal constraints? Recent empirical studies have revealed that experience with one set of causal cues can dramatically alter subsequent (...) class='Hi'>learning and performance with entirely different cues, suggesting that learning involves abstract transfer, and such transfer effects involve sequential presentation of distinct sets of causal cues. It has been demonstrated that pre-training can modulate classic causal learning phenomena such as forward and backward blocking. To account for these effects, we propose a Bayesian theory of sequential causal learning. The theory assumes that humans are able to consider and use several alternative causal generative models, each instantiating a different causal integration rule. Model selection is used to decide which integration rule to use in a given learning environment in order to infer causal knowledge from sequential data. Detailed computer simulations demonstrate that humans rely on the abstract characteristics of outcome variables to select a causal integration rule, which in turn alters causal learning in a variety of blocking and overshadowing paradigms. When the nature of the outcome variable is ambiguous, humans select the model that yields the best fit with the recent environment, and then apply it to subsequent learning tasks. Based on sequential patterns of cue-outcome co-occurrence, the theory can account for a range of phenomena in sequential causal learning, including various blocking effects, primacy effects in some experimental conditions, and apparently abstract transfer of causal knowledge. (shrink)

The ability to learn cause–effect relations from experience is critical for humans to behave adaptively — to choose causes that bring about desired effects. However, traditional experiments on experience-based learning involve events that are artificially compressed in time so that all learning occurs over the course of minutes. These paradigms therefore exclusively rely upon working memory. In contrast, in real-world situations we need to be able to learn cause–effect relations over days and weeks, which necessitates long-term memory. 413 (...) participants completed a smartphone study, which compared learning a cause–effect relation one trial per day for 24 days versus the traditional paradigm of 24 trials back- to- back. Surprisingly, we found few differences between the short versus long timeframes. Subjects were able to accurately detect generative and preventive causal relations, and they exhibited illusory correlations in both the short and long timeframe tasks. These results provide initial evidence that experience-based learning over long timeframes exhibits similar strengths and weaknesses as in short timeframes. However, learning over long timeframes may become more impaired with more complex tasks. (shrink)

Osiurak and Reynaud highlight the critical role of technical-reasoning skills in the emergence of human cumulative technological culture, in contrast to previous accounts foregrounding social-reasoning skills as key to CTC. We question their analysis of the available evidence, yet for other reasons applaud the emphasis on causal understanding as central to the adaptive and collaborative dynamics of CTC.

Learning to understand a single causal system can be an achievement, but humans must learn about multiple causal systems over the course of a lifetime. We present a hierarchical Bayesian framework that helps to explain how learning about several causal systems can accelerate learning about systems that are subsequently encountered. Given experience with a set of objects, our framework learns a causal model for each object and a causal schema that captures commonalities (...) among these causal models. The schema organizes the objects into categories and specifies the causal powers and characteristic features of these categories and the characteristic causal interactions between categories. A schema of this kind allows causal models for subsequent objects to be rapidly learned, and we explore this accelerated learning in four experiments. Our results confirm that humans learn rapidly about the causal powers of novel objects, and we show that our framework accounts better for our data than alternative models of causal learning. (shrink)

The field of causal learning and reasoning (largely overlooked in the target article) provides an illuminating case study of how the modern Bayesian framework has deepened theoretical understanding, resolved long-standing controversies, and guided development of new and more principled algorithmic models. This progress was guided in large part by the systematic formulation and empirical comparison of multiple alternative Bayesian models.

The power PC theory postulates a normative procedure for making causal inferences from contingency information, and offers this as a descriptive model of human causal judgement. The inferential procedure requires a set of assumptions, which includes the assumption that the cause being judged is distributed independently of the set of other possible causes of the same outcome. It is argued that this assumption either never holds or can never be known to hold. It is also argued that conformity (...) of judgements to the prescriptions of the model requires a sophisticated appreciation of methodological factors and acquired domain-specific knowledge of causes, and that the theory is disconfirmed by a finding that an objective contingency that equally supports two causal inferences results in only one of them actually being made. An alternative proposal based on the hypothesis that causal understanding originates with experiences of forces exerted while acting on objects is briefly sketched. (shrink)

Research on human causal learning has largely focused on strength learning, or on computational-level theories; there are few formal algorithmic models of how people learn causal structure from covariations. We introduce a model that learns causal structure in a local manner via prediction-error learning. This local learning is then integrated dynamically into a unified representation of causal structure. The model uses computationally plausible approximations of rational learning, and so represents a hybrid (...) between the associationist and rational paradigms in causal learning research. We conclude by showing that the model provides a good fit to data from a previous experiment. (shrink)

Real causal systems are complicated. Despite this, causal learning research has traditionally emphasized how causal relations can be induced on the basis of idealized events, i.e. those that have been mapped to binary variables and abstracted from time. For example, participants may be asked to assess the efficacy of a headache-relief pill on the basis of multiple patients who take the pill (or not) and find their headache relieved (or not). In contrast, the current study examines (...) learning via interactions with continuous dynamic systems, systems that include continuous variables that interact over time (and that can be continuously observed in real time by the learner). To explore such systems, we develop a new framework that represents a causal system as a network of stationary Gauss--Markov (``\OU'') processes and show how such \textit{OU networks} can express complex dynamic phenomena such as feedback loops and oscillations. To assess adult's abilities to learn such systems, we conducted an experiment in which participants were asked to identify the causal relationships of a number of OU networks, potentially carrying out multiple, temporally-extended interventions. We compared their judgments to a normative model for learning OU networks as well as a range of alternative and heuristic learning models from the literature. We found that, although participants exhibited substantial learning of such systems, they committed certain systematic errors. These successes and failures were best accounted for by a model that describes people as focusing on pairs of variables, rather than evaluating the evidence with respect to the full space of possible structural models. We argue that our approach provides both a principled framework for exploring the space of dynamic learning environments as well as new algorithmic insights into how people interact successfully with a continuous causal world. (shrink)

Causal structure learning algorithms have focused on learning in ”batch-mode”: i.e., when a full dataset is presented. In many domains, however, it is important to learn in an online fashion from sequential or ordered data, whether because of memory storage constraints or because of potential changes in the underlying causal structure over the course of learning. In this paper, we present TDSL, a novel causal structure learning algorithm that processes data sequentially. This algorithm (...) can track changes in the generating causal structure or parameters, and requires significantly less memory in realistic settings. We show by simulation that the algorithm performs comparably to batch-mode learning when the causal structure is stationary, and significantly better in non-stationary environments. (shrink)

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. (shrink)

Conditional learning, where agents learn a conditional sentence ‘If A, then B,’ is difficult to incorporate into existing Bayesian models of learning. This is because conditional learning is not uniform: in some cases, learning a conditional requires decreasing the probability of the antecedent, while in other cases, the antecedent probability stays constant or increases. I argue that how one learns a conditional depends on the causal structure relating the antecedent and the consequent, leading to a (...) causal model of conditional learning. This model extends traditional Bayesian learning by incorporating causal models into agents’ epistemic states. On this theory, conditional learning proceeds in two steps. First, an agent learns a new causal model with the appropriate relationship between the antecedent and the consequent. Then, the agent narrows down the set of possible worlds to include only those which make the conditional proposition true. This model of learning can incorporate both standard cases of Bayesian learning and the non-uniform learning required to learn conditional information. (shrink)

Machine learning is used more and more in scientific contexts, from the recent breakthroughs with AlphaFold2 in protein fold prediction to the use of ML in parametrization for large climate/astronomy models. Yet it is unclear whether we can obtain scientific explanations from such models. I argue that when machine learning is used to conduct causal inference we can give a new positive answer to this question. However, these ML models are purpose-built models and there are technical results (...) showing that standard machine learning models cannot be used for the same type of causal inference. Instead, there is a pathway to causal explanations from predictive ML models through new explainability techniques; specifically, new methods to extract structural equation models from such ML models. The extracted models are likely to suffer from issues though: they will often fail to account for confounders and colliders, as well as deliver simply incorrect causal graphs due to ML models tendency to violate physical laws such as the conservation of energy. In this case, extracted graphs are a starting point for new explanations, but predictive accuracy is no guarantee for good explanations. (shrink)

Discovering high-level causal relations from low-level data is an important and challenging problem that comes up frequently in the natural and social sciences. In a series of papers, Chalupka etal. (2015, 2016a, 2016b, 2017) develop a procedure forcausal feature learning (CFL) in an effortto automate this task. We argue that CFL does not recommend coarsening in cases where pragmatic considerations rule in favor of it, and recommends coarsening in cases where pragmatic considerations rule against it. We propose a (...) new technique, pragmatic causal feature learning (PCFL), which extends the original CFL algorithm in useful and intuitive ways. We show that PCFL has the same attractive measure-theoretic properties as the original CFL algorithm. We compare the performance of both methods through theoretical analysis and experiments. (shrink)

Consider the problem of learning, from non-experimental data, the causal (Markov equivalence) structure of the true, unknown causal Bayesian network (CBN) on a given, fixed set of (categorical) variables. This learning problem is known to be very hard, so much so that there is no learning algorithm that converges to the truth for all possible CBNs (on the given set of variables). So the convergence property has to be sacrificed for some CBNs—but for which? In (...) response, the standard practice has been to design and employ learning algorithms that secure the convergence property for at least all the CBNs that satisfy the famous faithfulness condition, which implies sacrificing the convergence property for some CBNs that violate the faithfulness condition (Spirtes, Glymour, and Scheines, 2000). This standard design practice can be justified by assuming—that is, accepting on faith—that the true, unknown CBN satisfies the faithfulness condition. But the real question is this: Is it possible to explain, without assuming the faithfulness condition or any of its weaker variants, why it is mandatory rather than optional to follow the standard design practice? This paper aims to answer the above question in the affirmative. We first define an array of modes of convergence to the truth as desiderata that might or might not be achieved by a causal learning algorithm. Those modes of convergence concern (i) how pervasive the domain of convergence is on the space of all possible CBNs and (ii) how uniformly the convergence happens. Then we prove a result to the following effect: for any learning algorithm that tackles the causal learning problem in question, if it achieves the best achievable mode of convergence (considered in this paper), then it must follow the standard design practice of converging to the truth for at least all CBNs that satisfy the faithfulness condition—it is a requirement, not an option. (shrink)

This paper defends the thesis of learning from non-causal models: viz. that the study of some model can prompt justified changes in one’s confidence in empirical hypotheses about a real-world target in the absence of any known or predicted similarity between model and target with regards to their causal features. Recognizing that we can learn from non-causal models matters not only to our understanding of past scientific achievements, but also to contemporary debates in the philosophy of (...) science. At one end of the philosophical spectrum, my thesis undermines the views of those who, like Cartwright (Erkenntnis 70:45–58, 2009), follow Hesse (Models and Analogies in Science, Notre Dame, University of Notre Dame Press, 1963) in restricting the possibility of learning from models to only those situations where a model identifies some causal factors present in the target. At the other end of the spectrum, my thesis also helps undermine some extremely permissive positions, e.g., Grüne-Yanoff’s (Erkenntnis 70(1):81–99, 2009, Philos Sci 80(5): 850–861, 2013) claim that learning from a model is possible even in the absence of any similarity at all between model and target. The thesis that we can learn from non-causal models offers a cautious middle ground between these two extremes. (shrink)

How ought we learn causal relationships? While Popper advocated a hypothetico-deductive logic of causal discovery, inductive accounts are currently in vogue. Many inductive approaches depend on the causal Markov condition as a fundamental assumption. This condition, I maintain, is not universally valid, though it is justifiable as a default assumption. In which case the results of the inductive causal learning procedure must be tested before they can be accepted. This yields a synthesis of the hypothetico-deductive (...) and inductive accounts, which forms the focus of this paper. I discuss the justification of this synthesis and draw an analogy between objective Bayesianism and the account of causal learning presented here. (shrink)

Introduction -- Simple linear causality : one thing makes another happen -- The cognitive science of simple causality : why do we get stuck? -- Domino causality : effects that become causes -- Cyclic causality : loops and feedback -- Spiraling causality : escalation and de-escalation -- Mutual causality : symbiosis and bi-directionality -- Relational causality : balances and differentials -- Across time and distance : detecting delayed and distant effects -- "What happened?" vs. "what's going on?" : thinking about (...) steady states -- What you can't see does matter : attending to obvious and non-obvious causes -- It's not always a case of who did it : minding passive and unintentional causality -- Step by step, or not : the mind-bending concept of simultaneous causality -- Figuring out what to count on : dealing with stochastic causality -- Isn't anybody in charge around here? : attending to distributed causality and emergence -- Summing up : the implications for helping a new generation understand causal complexity -- Putting it all together : teaching for causal complexity. (shrink)

Young children, in general, appear to have a strong drive to explore the environment in ways that reveal its underlying causal structure. But are they really attuned specifically to casual information in this quest for understanding, or do they show equal interest in other types of non-obvious information about the world? To answer this question, we introduced 20 three-year-old children to two puppets who were anxious to tell the child about a set of novel artifacts and animals. One puppet (...) consistently described causal properties of the items while the other puppet consistently described carefully matched non-causal properties of the same items. After a familiarization period in which children learned which type of information to expect from each informant, children were given the opportunity to choose which they wanted to hear describe each of eight pictured test items. On average, children chose to hear from the informant that provided causal descriptions on 72% of the trials. This preference for causal information has important implications for explaining the role of conceptual information in supporting early learning and may suggest means for maximizing interest and motivation in young children. (shrink)

Shane Harwood and Richard Scheines. Learning Linear Causal Structure Equation Models with Genetic Algorithms.

This work uses microgenetic study of classroom learning to illuminate (1) the role of pre-instructional student knowledge in the construction of normative scientific knowledge, and (2) the learning mechanisms that drive change. Three enactments of an instructional sequence designed to lead to a scientific understanding of thermal equilibration are used as data sources. Only data from a scaffolded student inquiry preceding introduction of a normative model were used. Hence, the study involves nearly autonomous student learning. In two (...) classes, students developed stable and socially shared explanations (“causal schemes”) for understanding thermal equilibration. One case resulted in a near-normative understanding, while the other resulted in a non-normative “alternative conception.” The near-normative case seems to be a particularly clear example wherein the constructed causal scheme is a composition of previously documented naïve conceptions. Detailed prior description of these naive elements allows a much better than usual view of the corresponding details of change during construction of the new scheme. A list of candidate mechanisms that can account for observed change is presented. The non-normative construction seems also to be a composition, albeit of a different structural form, using a different (although similar) set of naïve elements. This article provides one of very few high-resolution process analyses showing the productive use of naïve knowledge in learning. (shrink)

Can the phenomena of associative learning be replaced wholesale by a propositional reasoning system? Mitchell et al. make a strong case against an automatic, unconscious, and encapsulated associative system. However, their propositional account fails to distinguish inferences based on actions from those based on observation. Causal Bayes networks remedy this shortcoming, and also provide an overarching framework for both learning and reasoning. On this account, causal representations are primary, but associative learning processes are not excluded (...) a priori. (shrink)