Search

In this research, we investigated individual differences in the formation and persistence of causal illusions. In a re‐analysis of existing data, we identified two clusters of participants – persistent and adjusting – based on their trajectories in learning from repeated exposure to null contingencies. The persistent cluster maintained stable causal illusions, while the adjusting cluster demonstrated a reduction over time. This re‐analysis provided a nuanced understanding of individual differences in causal learning, emphasizing the differential role of probability estimations in predicting causal judgements. These findings were replicated in a subsequent study, highlighting the robustness of the identified effects. In a pre‐registered study, we extended the paradigm to include a second phase (active phase) to assess how individual differences in causal illusion trajectories in the passive phase would manifest when participants had agency in the information gathering process. The results were consistent with those of the two previous studies and confirmed our primary hypothesis that the adjusting cluster would exhibit a lower tendency to introduce the candidate cause on learning trials, and would, therefore, observe a higher frequency of cause–absent trials. Together, these studies provide comprehensive insights into the underpinnings of causal illusion development and persistence, potentially informing de‐biasing interventions. [ABSTRACT FROM AUTHOR]

Self-directed learners are described as "intuitive scientists," yet they often struggle in assessments of their scientific reasoning skills. We investigate a novel explanation for this apparent gap between formal and informal scientific inquiry behavior. Specifically, we consider whether learners' documented failure to correctly apply the control of variables strategy might stem from a mismatch between task presentation and their intuitions as causal learners. In Experiment 1, children (7- and 9-year-olds) and adults were tested on a version of a traditional multivariate reasoning task (Tschirgi, 1980) that was modified to clarify ambiguous elements of the causal logic in the original design. In all age groups, a significant majority of participants selected informative experiments on this modified task, avoiding confounded actions with positive tangible outcomes. In Experiment 2, we replicate these results with real-world stimuli, and in Experiments 3 and 4, we provide direct evidence that self-directed learners apply specific causal intuitions to experimentation tasks. Together, these findings support a novel alternative interpretation of the apparently paradoxical gap between learners' success in informal exploration and their error-prone experimentation—both behaviors are consistent with an intuitively causal approach to scientific inquiry. Public Significance Statement: From early life, we are precocious exploratory learners, and intrinsically motivated to seek informative evidence about the world. However, we also often struggle in contexts that require scientific experimentation. We provide evidence for a novel explanation for this disconnect: the standard control of variables tasks used to assess experimentation skills violates learners' intuitions about everyday causality. In a modified task, learners overwhelmingly select correct experiments, suggesting a novel continuity between exploration and scientific reasoning behavior. [ABSTRACT FROM AUTHOR]

Controllable Image Synthesis (CIS) is a methodology that allows users to generate desired images or manipulate specific attributes of images by providing precise input conditions or modifying latent representations. In recent years, CIS has attracted considerable attention in the field of image processing, with significant advances in consistency, controllability and harmony. However, several challenges still remain, particularly regarding the fine-grained controllability and interpretability of synthesized images. In this paper, we comprehensively and systematically review the CIS from problem definition, taxonomy and evaluation systems to existing challenges and future research directions. First, the definition of CIS is given, and several representative deep generative models are introduced in detail. Second, the existing CIS methods are divided into three categories according to the different control manners used and discuss the typical work in each category critically. Furthermore, we introduce the public datasets and evaluation metrics commonly used in image synthesis and analyze the representative CIS methods. Finally, we present several open issues and discuss the future research direction of CIS. [ABSTRACT FROM AUTHOR]

Rice consistently faces significant threats from biotic stresses, such as fungi, bacteria, pests, and viruses. Consequently, accurately and rapidly identifying previously unknown single-nucleotide polymorphisms (SNPs) in the rice genome is a critical challenge for rice research and the development of resistant varieties. However, the limited availability of high-quality rice genotype data has hindered this research. Deep learning has transformed biological research by facilitating the prediction and analysis of SNPs in biological sequence data. Convolutional neural networks are especially effective in extracting structural and local features from DNA sequences, leading to significant advancements in genomics. Nevertheless, the expanding catalog of genome-wide association studies provides valuable biological insights for rice research. Expanding on this idea, we introduce RiceSNP-BST, an automatic architecture search framework designed to predict SNPs associated with rice biotic stress traits (BST-associated SNPs) by integrating multidimensional features. Notably, the model successfully innovates the datasets, offering more precision than state-of-the-art methods while demonstrating good performance on an independent test set and cross-species datasets. Additionally, we extracted features from the original DNA sequences and employed causal inference to enhance the biological interpretability of the model. This study highlights the potential of RiceSNP-BST in advancing genome prediction in rice. Furthermore, a user-friendly web server for RiceSNP-BST (http://rice-snp-bst.aielab.cc) has been developed to support broader genome research. [ABSTRACT FROM AUTHOR]

The notion of causality assumes a paramount position within the realm of human cognition. Over the past few decades, there has been significant advancement in the domain of causal effect estimation across various disciplines, including but not limited to computer science, medicine, economics, and industrial applications. Given the continous advancements in deep learning methodologies, there has been a notable surge in its utilization for the estimation of causal effects using counterfactual data. Typically, deep causal models map the characteristics of covariates to a representation space and then design various objective functions to estimate counterfactual data unbiasedly. Different from the existing surveys on causal models in machine learning, this review mainly focuses on the overview of the deep causal models based on neural networks, and its core contributions are as follows: (1) we cast insight on a comprehensive overview of deep causal models from both timeline of development and method classification perspectives; (2) we outline some typical applications of causal effect estimation to industry; (3) we also endeavor to present a detailed categorization and analysis on relevant datasets, source codes and experiments. [ABSTRACT FROM AUTHOR]

Much discussion about large language models and language-and-vision models has focused on whether these models are intelligent agents. We present an alternative perspective. First, we argue that these artificial intelligence (AI) models are cultural technologies that enhance cultural transmission and are efficient and powerful imitation engines. Second, we explore what AI models can tell us about imitation and innovation by testing whether they can be used to discover new tools and novel causal structures and contrasting their responses with those of human children. Our work serves as a first step in determining which particular representations and competences, as well as which kinds of knowledge or skill, can be derived from particular learning techniques and data. In particular, we explore which kinds of cognitive capacities can be enabled by statistical analysis of large-scale linguistic data. Critically, our findings suggest that machines may need more than large-scale language and image data to allow the kinds of innovation that a small child can produce. [ABSTRACT FROM AUTHOR]

People often rely on the covariation between events to infer causality. However, covariation between cues and outcomes may change over time. In the associative learning literature, extinction provides a model to study updating of causal beliefs when a previously established relationship no longer holds. Prediction error theories can explain both extinction and protection from extinction when an inhibitory (preventive) cue is present during extinction. In three experiments using the allergist causal learning task, we found that protection could also be achieved by a hidden cause that was inferred but not physically present, so long as that cause was a plausible preventer of the outcome. We additionally showed complete protection by a physically presented cue that was neutral rather than inhibitory at the outset of extinction. Both findings are difficult to reconcile with dominant prediction error theories. However, they are compatible with the idea of theory protection, where the learner attributes the absence of the outcome to the added cue (when present) or to a hidden cause, and therefore does not need to revise their causal beliefs. Our results suggest that prediction error encourages changes in causal beliefs, but the nature of the change is determined by reasoning processes that incorporate existing knowledge of causal mechanisms and may be biased toward preservation of existing beliefs. [ABSTRACT FROM AUTHOR]

Causal reasoning—the ability to reason about causal relations between events—is fundamental to understanding how the world works. This paper reviews two prominent theories on early causal learning and offers possibilities for theory bridging. Both theories grow out of computational modeling and have significant areas of overlap while differing in several respects. Explanation‐Based Learning (EBL) focuses on young infants' learning about causal concepts of physical objects and events, whereas Bayesian models have been used to describe causal reasoning beyond infancy across various concept domains. Connecting the two models offers a more integrated approach to clarifying the developmental processes in causal reasoning from early infancy through later childhood. We further suggest that everyday language practices offer a promising space for theory bridging. We provide a review of selective work on caregiver–child conversations, in particular, on the use of scaffolding language including causal talk and pedagogical questions. Linking the research on language practices to the two cognitive theories, we point out directions for further research to integrate EBL and Bayesian models and clarify how causal learning unfolds in real life. This article is categorized under:Psychology > LearningCognitive Biology > Cognitive Development [ABSTRACT FROM AUTHOR]

Although facial expression recognition based on deep learning has become a major trend, existing methods have been found to prefer learning spurious statistical correlations and non-robust features during training. This degenerates the model's generalizability in practical situations. One of the research fields mitigating such misperception of correlations as causality is causal reasoning. In this article, we propose a learnable counterfactual attention mechanism, CA-FER, that uses causal reasoning to simultaneously optimize feature discrimination and diversity to mitigate spurious correlations in expression datasets. To the best of our knowledge, this is the first work to study the spurious correlations in facial expression recognition from a counterfactual attention perspective. Extensive experiments on a synthetic dataset and four public datasets demonstrate that our method outperforms previous methods, which shows the effectiveness and generalizability of our learnable counterfactual attention mechanism for the expression recognition task. [ABSTRACT FROM AUTHOR]

Investigating causality to establish novel criteria for training robust natural language processing (NLP) models is an active research area. However, current methods face various challenges such as the difficulties in identifying keyword lexicons and obtaining data from multiple labeled environments. In this paper, we study the problem of robust NLP from a complementary but different angle: we treat the behavior of an attack model as a complex causal mechanism and quantify its algorithmic information using the minimum description length (MDL) framework. Specifically, we use masked language modeling (MLM) to measure the "amount of effort" needed to transform from the original text to the altered text. Based on that, we develop techniques for judging whether a specified set of tokens has been altered by the attack, even in the absence of the original text data. [ABSTRACT FROM AUTHOR]

Multimodal remote sensing data classification can enhance a model's ability to distinguish land features through multimodal data fusion. In this context, how to help models understand the relationship between multimodal data and target tasks has become the focus of researchers. Inspired by the human feedback learning mechanism, causal reasoning mechanism, and knowledge induction mechanism, this paper integrates causal learning, reinforcement learning, and meta learning into a unified remote sensing data classification framework and proposes causal meta-reinforcement learning (CMRL). First, based on the feedback learning mechanism, we overcame the limitations of traditional implicit optimization of fusion features and customized a reinforcement learning environment for multimodal remote sensing data classification tasks. Through feedback interactive learning between agents and the environment, we helped the agents understand the complex relationships between multimodal data and labels, thereby achieving full mining of multimodal complementary information.Second, based on the causal inference mechanism, we designed causal distribution prediction actions, classification rewards, and causal intervention rewards, capturing pure causal factors in multimodal data and preventing false statistical associations between non-causal factors and class labels. Finally, based on the knowledge induction mechanism, we designed a bi-layer optimization mechanism based on meta-learning. By constructing a meta training task and meta validation task simulation model in the generalization scenario of unseen data, we helped the model induce cross-task shared knowledge, thereby improving its generalization ability for unseen multimodal data. The experimental results on multiple sets of multimodal datasets showed that the proposed method achieved state-of-the-art performance in multimodal remote sensing data classification tasks. [ABSTRACT FROM AUTHOR]

The outcome of any scientific experiment or intervention will naturally unfold over time. How then should individuals make causal inferences from measurements over time? Across three experiments, we had participants observe experimental and control groups over several days posttreatment in a fictional biological research setting. We identify competing perspectives in the literature: contingency-driven accounts predict no effect of the outcome timing while the contiguity principle suggests people will view a treatment as more harmful to the extent that bad treatment outcomes occur earlier rather than later. In contrast, inference of the functional form of a treatment effect can license extrapolation beyond the measurements and lead to different causal inferences. We find participants' causal strength and direction judgments in temporal settings vary with minimal manipulations of instruction framing. When it is implied that the observations are made over a preplanned number of days, causal judgments depend strongly on contiguity. When it is implied that the observation may be ongoing, participants extrapolate current trends into the future and adapt their causal judgments accordingly. When data are revealed sequentially, participants rely on extrapolation regardless of instruction framing. Our results demonstrate human flexibility in interpreting temporal evidence for causal reasoning and emphasize human tendency to generalize from evidence in ways that are acutely sensitive to task framing. [ABSTRACT FROM AUTHOR]

Illusory causation is a phenomenon in which people mistakenly perceive a causal relationship between a cue and outcome even though the contingency between them is actually zero. Illusory causation studies typically use a unidirectional causal rating scale, where one endpoint refers to no relationship and the other to a strongly positive causal relationship. This procedure may bias mean causal ratings in a positive direction, either by censoring negative ratings or by discouraging participants from giving the normative rating of zero which is at the bottom extreme of the scale. To test this possibility, we ran two experiments that directly compared the magnitude of causal illusions when assessed with a unidirectional (zero—positive) versus a bidirectional (negative—zero—positive) rating scale. Experiment 1 used high cue and outcome densities (both 75%), whereas Experiment 2 used neutral cue and outcome densities (both 50%). Across both experiments, we observed a larger illusory causation effect in the unidirectional group compared with the bidirectional group, despite both groups experiencing the same training trials. The causal illusions in Experiment 2 were observed despite participants accurately learning the conditional probabilities of the outcome occurring in both the presence and absence of the cue, suggesting that the illusion is driven by the inability to accurately integrate conditional probabilities to infer causal relationships. Our results indicate that although illusory causation is a genuine phenomenon that is observable with either a undirectional or a bidirectional rating scale, its magnitude may be overestimated when unidirectional rating scales are used. [ABSTRACT FROM AUTHOR]

Considerable delays between causes and effects are commonly found in real life. However, previous studies have only investigated how well people can learn probabilistic relations with delays on the order of seconds. In the current study we tested whether people can learn a cause-effect relation with delays of 0, 3, 9, or 21 hours, and the study lasted 16 days. We found that learning was slowed with longer delays, but by the end of 16 days participants had learned the cause-effect relation in all four conditions, and they had learned the relation about equally well in all four conditions. This suggests that in real-world situations people may still be fairly accurate at inferring cause-effect relations with delays if they have enough experience. We also discuss ways that delays may interact with other real-world factors that could complicate learning. [ABSTRACT FROM AUTHOR]

Biometric signal based human-computer interface (HCI) has attracted increasing attention due to its wide application in healthcare, entertainment, neurocomputing, and so on. In recent years, deep learning-based approaches have made great progress on biometric signal processing. However, the state-of-the-art (SOTA) approaches still suffer from model degradation across subjects or sessions. In this work, we propose a novel unsupervised domain adaptation approach for biometric signal-based HCI via causal representation learning. Specifically, three kinds of interventions on biometric signals (i.e., subjects, sessions, and trials) can be selected to generalize deep models across the selected intervention. In the proposed approach, a generative model is trained for producing intervened features that are subsequently used for learning transferable and causal relations with three modes. Experiments on the EEG-based emotion recognition task and sEMG-based gesture recognition task are conducted to confirm the superiority of our approach. An improvement of +0.21% on the task of inter-subject EEG-based emotion recognition is achieved using our approach. Besides, on the task of inter-session sEMG-based gesture recognition, our approach achieves improvements of +1.47%, +3.36%, +1.71%, and +1.01% on sEMG datasets including CSL-HDEMG, CapgMyo DB-b, 3DC, and Ninapro DB6, respectively. The proposed approach also works on the task of inter-trial sEMG-based gesture recognition and an average improvement of +0.66% on Ninapro databases is achieved. These experimental results show the superiority of the proposed approach compared with the SOTA unsupervised domain adaptation methods on HCIs based on biometric signal. [ABSTRACT FROM AUTHOR]

In multi-label learning, the label-specific features learning framework can effectively solve the dimensional catastrophe problem brought by high-dimensional data. The classification performance and robustness of the model are effectively improved. Most existing label-specific features learning utilizes the cosine similarity method to measure label correlation. It is well known that the correlation between labels is asymmetric. However, existing label-specific features learning only considers the private features of labels in classification and does not take into account the common features of labels. Based on this, this paper proposes a Causality-driven Common and Label-specific Features Learning, named CCSF algorithm. Firstly, the causal learning algorithm GSBN is used to calculate the asymmetric correlation between labels. Then, in the optimization, both -norm and -norm are used to select the corresponding features, respectively. Finally, it is compared with six state-of-the-art algorithms on nine datasets. The experimental results prove the effectiveness of the algorithm in this paper. [ABSTRACT FROM AUTHOR]

The present paper examines a type of abstract domain-general knowledge required for the process of constructing useable domain-specific causal knowledge, the evident goal of causal learning. It tests the hypothesis that analytic knowledge of causal-invariance decomposition functions is essential for this process. Such knowledge specifies the decomposition of an observed outcome into contributions from constituent causes under the default assumption that the empirical knowledge acquired is invariant across contextual/background causes. The paper reports two psychological experiments (and replication studies) with pre-school-age children on generalization across contexts involving binary cause and effect variables. The critical role of causal invariance for constructing useable causal knowledge predicts that even young children should (tacitly) use the causal-invariance decomposition function for such variables rather than a non-causal-invariance decomposition function common in statistical practice in research involving binary outcomes. The findings support the rational shaping of empirical causal knowledge by the causal-invariance constraint, ruling out alternative explanations in terms of non-causal-invariance decomposition functions, heuristics, and biases. For the same causal structure involving candidate causes and outcomes that are binary variables with a "present" value and an "absent" value, the paper argues against the possibility of multiple rational characterizations of the "sameness of causal influence" that justifies generalization across contexts.

To improve upon their current knowledge, learners must be able to generate informative data and accurately evaluate this evidence. However, there is substantial disagreement regarding self-directed learners' competence in these behaviors. Researchers in cognitive development have suggested that learners are "intuitive scientists," generating informative actions and rationally coordinating their current observations and prior beliefs from an early age. Conversely, researchers in scientific reasoning report that learners struggle with experimentation and often fail to reach appropriate conclusions from evidence, even as adults. According to the prevailing narrative, these inconsistent findings must be "bridged" to explain the gap between learners' successes and failures. Here, we advocate for an alternative approach. First, we review the research on scientific reasoning and find that there may be less evidence for learners' failures than is typically assumed. Second, we offer a novel interpretation that aims to account for both literatures: we suggest that self-directed learners may be best understood as competent causal reasoners. That is, many seemingly uninformative or irrational behaviors are consistent with the goals of causal learning. This account not only resolves the apparent contradictions in the existing research, but also offers a way forward towards a more accurate and integrated understanding of self-directed learning. This article is categorized under: Psychology > Development and Aging Psychology > Learning Psychology > Reasoning and Decision Making.

The article focuses on enhancing recommender systems by addressing the challenges of mixed causal and non-causal associations in user feedback. Topics include the proposal of a model-agnostic causal learning framework called IV4Rec that uses search queries as instrumental variables, the decomposition of embedding vectors into causal and non-causal parts, and the improvement of recommendation system performance through this approach.

Investigating causality to establish novel criteria for training robust natural language processing (NLP) models is an active research area. However, current methods face various challenges such as the difficulties in identifying keyword lexicons and obtaining data from multiple labeled environments. In this paper, we study the problem of robust NLP from a complementary but different angle: we treat the behavior of an attack model as a complex causal mechanism and quantify its algorithmic information using the minimum description length (MDL) framework. Specifically, we use masked language modeling (MLM) to measure the “amount of effort” needed to transform from the original text to the altered text. Based on that, we develop techniques for judging whether a specified set of tokens has been altered by the attack, even in the absence of the original text data.

Few-Shot Text Classification (FSTC) is a fundamental natural language processing problem that aims to classify small amounts of text with high accuracy. Mainstream methods model the superficial statistical relationships between text and labels. However, distributional imbalance problems are encountered during few-shot learning; therefore, questions remain regarding its robustness and generalization. The above problems can be addressed by intrinsic causal mechanisms. We introduce a general structural causal model to formalize the FSTC problem. To extract causal associations from text and reconstruct information to achieve a better classification effect, we propose a causal representation for few-shot learning (CRFL) framework to force representations to be causally related. Our framework performs well when the number of training examples is small or when it generalizes to the data transfer situation. CRFL is orthogonal to the existing fine-tuning and few-shot meta-learning methods and can be applied to any task. Extensive experimental results obtained on several widely used datasets validate the effectiveness of our approach, which can be attributed to our model's stability and logical reasoning. [ABSTRACT FROM AUTHOR]

A Random Graph is a random object which takes its values in a space of graphs. We take advantage of the expressibility of graphs in order to model uncertainty about the existence of causal relations within a given set of variables. We adopt a Bayesian point of view which leads us to propose a belief updating procedure with the objective of capturing a causal structure via interaction with a causal environment. Besides learning a causal structure, our proposal is also able to learn optimal actions, i.e., an optimal policy. We test our method in two experiments, each on a different scenario. Our experiments confirm that the proposed technique is able to learn a causal structure as well as an optimal policy. On the other hand, the second experiment shows that our proposal manages to learn an underlying causal model within several tasks in which different configurations of the causal structure are used. [ABSTRACT FROM AUTHOR]

Understanding that hypothesis testing is aimed at seeking information rather than producing desirable outcomes is indispensable for designing informative experiments. This study investigated the developmental course of information seeking compared to producing an effect in young children. In a between-subjects design, 4- to 6-year-olds (N = 109) were presented with the same pattern of events and asked to choose an intervention to either seek information on whether a hypothesis is right or produce an effect. The results revealed a developmental change between 4 and 6 years of age. From 5.5 years and onwards, children selectively chose the correct interventions in both conditions, whereas 4- and younger 5-year-olds did not. Contrary to the general view that producing desirable outcomes developmentally precedes engaging in information seeking, the present findings indicate that they follow a similar developmental line in the early childhood years. [ABSTRACT FROM AUTHOR]

We investigated whether children preferentially select informative actions and make accurate inferences from the outcome of their own interventions in a causal learning task. Four- to six-year-olds were presented with a novel system composed of gears that could operate according to two possible causal structures (single or multiple cause). Given the choice between interventions (i.e., removing one of the two gears to observe the remaining gear in isolation), children demonstrated a clear preference for the action that revealed the true causal structure, and made subsequent causal judgments that were consistent with the outcome observed. Experiment 2 addressed the possibility that performance was driven by children's tendency to select an intervention that would produce a desirable effect (i.e., spinning gears), rather than to disambiguate the causal structure. These results replicate our initial findings in a context in which the informative action was less likely to produce a positive outcome than the uninformative one. Experiment 3 serves as a control demonstrating that children's success in the previous experiments is not due to their use of low-level strategies. We discuss these findings in terms of their significance for understanding the development of scientific reasoning and the role of self-directed actions in early causal learning.

Recent work has shown that people use temporal informationincluding order, delay, and variability to infer causality be-tween events. In this study we build on this work by investi-gating the role of time in dynamic systems, where causes takecontinuous values and also continually influence their effects.Recent studies of learning in these systems explored short in-teractions in a setting with comparatively rapidly evolving dy-namics and modeled people as relying on simpler, resource-limited strategies to grapple with the stream of information(Davis et al., 2020). A natural question that arises from such anaccount is whether interacting with systems that unfold moreslowly might reduce the systematic errors that result from thesestrategies. Paradoxically, we find that slowing the task indeedreduced the frequency of one type of error, but increased the er-ror rate overall. To capture the differences between conditions,we introduce a novel Causal Event Segmentation model basedon the notion that people compress the continuous scenes intoevents and use these to drive structure inference.

When making causal inferences, prior research shows that peopleare capable of controlling for alternative causes. These studies,however, utilize artificial inter-trial intervals on the order ofseconds; in real-life situations people often experience data overdays and weeks (e.g., learning the effectiveness of two newmedications over multiple weeks). In the current study, participantslearned about two possible causes from data presented in atraditional trial-by-trial paradigm (rapid series of trials) versus amore naturalistic paradigm (one trial per day for multiple weeks viasmartphone). Our results suggest that while people are capable ofdetecting simple cause-effect relations that do not requirecontrolling for another cause when learning over weeks, they havedifficulty learning cause-effect relations that require controlling foralternative causes.

A large literature on the development of causal reasoningcharacterizes early childhood as a period of curiosity,exploration, and experimentation. This suggests that a noveltypreference may be a universal hallmark of early causallearning. Functionally, such a bias might serve to directattention towards new opportunities for knowledge gain. Analternative possibility is that a preference for exploring noveloutcomes develops over time. In three experiments with 2- to5-year-olds, we investigate the developmental trajectory ofchildren’s preference for causal processes that producereliable versus novel outcomes. We find evidence for adevelopmental shift between ages 2 and 3: while two-year-olds trend toward a preference for reliable over noveloutcomes, older children clearly prefer novel ones. Wediscuss possible adaptive reasons for this developmental shift.

How do people use temporal information to make causal judg-ments? A number of studies have investigated the role of timein inferring generative causal structure, while few have exam-ined prevention. Here, we focus on a challenging task in whichparticipants learn the structure of several causal “devices” bywatching the devices’ patterns of activation over time. Eachdevice potentially includes both generative (producing an acti-vation of its effect) and preventative (blocking any effect acti-vations within a short time window) causal relationships. Weexamine judgment patterns through the lens of a normativemodel which incorporates actual causation with considerationsof prevention. We contrast this with a more computationallytractable feature-based approximation. Participants’ perfor-mance was substantially above chance in all conditions. Themajority of participants’ causal judgments were best fit by thefeature-based approximation based on delay and count heuris-tic cues.

Causal inference can estimate causal effects, but unless data are collected experimentally, statistical analyses must rely on pre-specified causal models. Causal discovery algorithms are empirical methods for constructing such causal models from data. Several asymptotically correct discovery methods already exist, but they generally struggle on smaller samples. Moreover, most methods focus on very sparse causal models, which may not always be a realistic representation of real-life data generating mechanisms. Finally, while causal relationships suggested by the methods often hold true, their claims about causal non-relatedness have high error rates. This non-conservative error trade off is not ideal for observational sciences, where the resulting model is directly used to inform causal inference: A causal model with many missing causal relations entails too strong assumptions and may lead to biased effect estimates. We propose a new causal discovery method that addresses these three shortcomings: Supervised learning discovery (SLdisco). SLdisco uses supervised machine learning to obtain a mapping from observational data to equivalence classes of causal models. We evaluate SLdisco in a large simulation study based on Gaussian data and we consider several choices of model size and sample size. We find that SLdisco is more conservative, only moderately less informative and less sensitive towards sample size than existing procedures. We furthermore provide a real epidemiological data application. We use random subsampling to investigate real data performance on small samples and again find that SLdisco is less sensitive towards sample size and hence seems to better utilize the information available in small datasets. [ABSTRACT FROM AUTHOR]

Multimodal remote sensing data classification can enhance a model’s ability to distinguish land features through multimodal data fusion. In this context, how to help models understand the relationship between multimodal data and target tasks has become the focus of researchers. Inspired by the human feedback learning mechanism, causal reasoning mechanism, and knowledge induction mechanism, this paper integrates causal learning, reinforcement learning, and meta learning into a unified remote sensing data classification framework and proposes causal meta-reinforcement learning (CMRL). First, based on the feedback learning mechanism, we overcame the limitations of traditional implicit optimization of fusion features and customized a reinforcement learning environment for multimodal remote sensing data classification tasks. Through feedback interactive learning between agents and the environment, we helped the agents understand the complex relationships between multimodal data and labels, thereby achieving full mining of multimodal complementary information.Second, based on the causal inference mechanism, we designed causal distribution prediction actions, classification rewards, and causal intervention rewards, capturing pure causal factors in multimodal data and preventing false statistical associations between non-causal factors and class labels. Finally, based on the knowledge induction mechanism, we designed a bi-layer optimization mechanism based on meta-learning. By constructing a meta training task and meta validation task simulation model in the generalization scenario of unseen data, we helped the model induce cross-task shared knowledge, thereby improving its generalization ability for unseen multimodal data. The experimental results on multiple sets of multimodal datasets showed that the proposed method achieved state-of-the-art performance in multimodal remote sensing data classification tasks.

How do children’s beliefs about a causal system influence theirexploration of that system? Children watched an experimentertry to make a machine play music by placing blocks on top; oneblock always activated the machine and the other block neverdid (Deterministic condition), or one block activated the ma-chine a higher proportion of times than the other (Probabilisticcondition). Subsequently, we measured children’s exploratorybehaviors without feedback (the machine never activated). Wepredicted that children in the two conditions would differ intheir beliefs about how the system should work, leading to dif-ferent hypotheses about why the machine was no longer work-ing, and to differential exploration. Compared to the Proba-bilistic condition, children in the Deterministic condition in-tervened more often with the previously more effective block,experimented more with how to activate the machine, and ex-plored for less time. Children’s exploration provides a rich,nuanced view of their causal reasoning.

We investigate whether children preferentially selectinformative actions and make accurate inferences from theoutcome of their own interventions in a causal learning task.Four- to six-year-olds were presented with a novel systemcomposed of two gears that could operate according to twopossible causal structures (single or multiple cause). Giventhe choice between interventions (i.e., removing one of thegears to observe the remaining gear in isolation), childrendemonstrated a clear preference for the action that revealedthe true causal structure, and made subsequent causaljudgments that were consistent with the outcome observed.Experiment 2 addressed the possibility that performance wasdriven by children’s tendency to select an intervention thatwould produce a desirable effect (i.e., spinning gears), ratherthan to disambiguate the causal structure. The results replicateour initial findings in a context in which the informativeaction was less likely to produce a positive outcome than theuninformative one. We discuss these results in terms of theirsignificance for understanding both the development ofscientific reasoning and the role of self-directed actions inearly learning.

Intervention selection is at once crucial in causal learning andchallenging for causal learners. While the optimal strategy ismaximizing the expected information gain (EIG), both chil-dren and adults often combine it with suboptimal ones suchas the positive test strategy (PTS). In the current study, wesought to facilitate causal learning from intervention by asking5- to 7-year-olds to explain why they chose a certain interven-tion to identify the true structure of a three-node causal sys-tem that might work in one of two ways. Our findings suggestthat while engaging in self-explaining did not help children se-lect more informative interventions, asking them to think abouttheir intervention choices (explaining or reporting) might helpthem better utilize interventional data to infer causal structures.

Intervening on causal systems can illuminate their underlyingstructures. Past work has shown that, relative to adults, youngchildren often make intervention decisions that confirm sin-gle hypotheses rather than those that discriminate alternativehypotheses. Here, we investigated how the ability to make in-formative intervention decisions changes across development.Ninety participants between the ages of 7 and 25 completed40 different puzzles in which they had to intervene on vari-ous causal systems to determine their underlying structures.We found that the use of discriminatory strategies increasedthrough adolescence and plateaued into adulthood. Our resultsidentify a clear developmental trend in causal reasoning, andhighlight the need to expand research on causal learning mech-anisms in adolescence.

Children often fail to control variables when conducting testsof hypotheses, yielding confounded evidence. We propose thatgetting children to think of alternative possibilities throughcounterfactual prompts may scaffold their ability to controlvariables, by engaging them in an imagined intervention that isstructurally similar to controlled actions in scientificexperiments. Findings provide preliminary support for thishypothesis. Seven- to 10-year-olds who were prompted to thinkcounterfactually showed better performance on post-testcontrol of variables tasks than children who were given controlprompts. These results inform debates about the contributionof counterfactual reasoning to scientific reasoning, and suggestthat counterfactual prompts may be useful in science learningcontexts.

Humans often rely on past experiences stored in long-termmemory to predict the outcome of an event. In traditional lab-based experiments (e.g., causal learning, probability learning,etc.), these observations are compressed into a successiveseries of learning trials. The rapid nature of this paradigmmeans that completing the task relies on working memory. Incontrast, real-world events are typically spread out over longerperiods of time, and therefore long-term memory must be used.We conducted a 24 day smartphone study to assess how wellpeople can learn causal relationships in extended timeframes.Surprisingly, we found few differences in causal learning whensubjects observed events in a traditional rapid series of 24 trialsas opposed to one trial per day for 24 days. Specifically,subjects were able to detect causality for generative andpreventive datasets and also exhibited illusory correlations inboth the short-term and long-term designs. We discusstheoretical implications of this work.

Children's understanding of contagion has been a fruitful area for studying children's learning. However, despite a large literature on children's conceptual understanding of illness, there is very little research on the impact of children's knowledge about illness transmission on adaptive behavior. This is important because how children behave when faced with a sick individual or a contaminated object is what is most relevant to whether children get sick and pass along that illness to other people. Here, we will bring together various theories of how children learn to behave adaptively when faced with the possibility of getting sick (a) to better illuminate the different ways by which children might acquire health-related behaviors and (b) to help develop recommendations for designing interventions aimed at teaching children about contagion and illness prevention in a way that produces the most adaptive health behaviors. [ABSTRACT FROM AUTHOR]

Three experiments investigated whether the nature of the temporal referent affects timing behaviour in rats. We used a peak procedure and assessed timing of food well activity as a function of whether the referent was an instrumental response (a lever press that resulted in the withdrawal of the lever) or a conditioned stimulus (CS) that was 2 s in Experiment 1, 500 ms in Experiment 2, and 800 ms in Experiment 3. In all experiments, the interval between the offset of the temporal referent and food was 5 s. The curve fits for each experiment revealed no differences in peak time, but magazine responding immediately following the CS was higher than following a lever press. This pattern of results was interpreted as reflecting a combination of (a) ambiguity in which component of the 500 ms–2-s auditory stimulus was serving as the referent and (b) response competition between lever pressing and magazine activity. Critically, these results suggest that peak timing in rats is unaffected by whether a lever press or CS serves as the referent. This conclusion is consistent with theoretical models of timing behaviour, but not with evidence from humans showing that the subjective perception of time is affected by whether the cause of an outcome was self-generated or not. [ABSTRACT FROM AUTHOR]

Learning invariant (causal) features for out-of-distribution (OOD) generalization have attracted extensive attention recently, and among the proposals, invariant risk minimization (IRM) is a notable solution. In spite of its theoretical promise for linear regression, the challenges of using IRM in linear classification problems remain. By introducing the information bottleneck (IB) principle into the learning of IRM, the IB-IRM approach has demonstrated its power to solve these challenges. In this paper, we further improve IB-IRM from two aspects. First, we show that the key assumption of support overlap of invariant features used in IB-IRM guarantees OOD generalization, and it is still possible to achieve the optimal solution without this assumption. Second, we illustrate two failure modes where IB-IRM (and IRM) could fail in learning the invariant features, and to address such failures, we propose a Counterfactual Supervision-based Information Bottleneck (CSIB) learning algorithm that recovers the invariant features. By requiring counterfactual inference, CSIB works even when accessing data from a single environment. Empirical experiments on several datasets verify our theoretical results. [ABSTRACT FROM AUTHOR]