United States

Prediction, explanation, and control are basic cognitive tasks. Here we show how they can arise from unstructured exploration-based learning. We present a new method to analyze how people make use of mental models learned under free exploration. Our method allows us to control for the relative difficulty of the task, and to measure the extent to which participants can leverage the evidence provided by the mental model---however strong or weak---for decision-making. The key result of our subsequent experimental work is that free exploration leads, in simple cases, to far better performance on prediction and control compared to explanation. As systems become more complex, performance declines, but people can sustain relatively better performance on control. In the presence of hidden variables, explanatory abilities rise relative to the other two, even though the task is harder.


CogSci 2023

Prediction, Explanation, and Control Under Free Exploration

poster

The 45th Annual Meeting of the Cognitive Science Society will take place in Sydney, Australia as a modified hybrid conference. Presentations from the main-session room will be live-streamed every day for all events. This will include all invited talks and award sessions. Recordings of all CogSci 2023 sessions will be made available for on-demand viewing after the conference.<br>
The in-person component will be held at the International Convention Centre Sydney at Darling Harbour. The Conference will take place Wednesday July 26th – Saturday July 29th, 2023 (Wednesday, July 26th is the Workshops Day).

Join us!

<br>[**The Program at a Glance can be found here**](https://cognitivesciencesociety.org/wp-content/uploads/2023/07/COgSci-PAAG-V2.pdf)
<br>[**The Detailed Program can be found here**](https://online.fliphtml5.com/ebtyf/mswz/)
<br><br>

#### **CogSci 2023 Program Book**
<iframe style="width:700px;height:400px" src="https://online.fliphtml5.com/ebtyf/tzgk/"  seamless="seamless" scrolling="no" frameborder="0" allowtransparency="true" allowfullscreen="true" ></iframe>
<br>

The 45th Annual Meeting of the Cognitive Science Society presents the latest research across cognitive science and highlights the theme of Cognition in Context.

We explore intervention policies chosen by participants when learning about causal structures in continuous systems. We used the continuous time and state space formalism of Ornstein-Uhlenbeck networks (Davis et al., 2020b). We find that participants' interventions can be understood as compositions of two basic actions: swipes, large movements generating large changes in values in a short time window, and holds, fixing variables for a few frames and observing the outcome. We further show that intervention policies depend on the underlying causal structure that participants are attempting to retrieve. Swipes were performed more often when there was no outgoing link, suggesting that rapid successive movements was useful in validating the absence of causal relations. On the other hand, interventions featuring longer holding times were favoured in cases where there was at least one outgoing link from the intervened node and appeared to serve the purpose of establishing the sign and strength of causal relationships. Overall, these findings indicate that causal inference in continuous contexts appears to be a two step process, where participants use different interventions for causal discovery than they do to qualify the nature of discovered links, hinting at a tendency to decompose inference to simplify it.

Swipe and hold: composing interventions in continuous time causal learning

Recent work suggests that decisions to detain defendants before trial increase the likelihood of conviction. One reason may be that knowledge of detainment makes jurors more likely to convict. Previous work has claimed this as an example of ‘bounded rationality’ i.e., due to a simple bias. We argue that this inference represents sophisticated causal reasoning e.g. about information hidden from jurors such as criminal history. We examine whether the effect of detainment knowledge on conviction depends on rational inference by presenting participants with a legal vignette in a 2x2 design: a defendant either has or has not been detained, and this detainment decision is either (1) not explained or (2) explained as due to an iron clad rule always used for this class of crimes. We find an effect of detainment when it is not explained, but no effect when explained, providing evidence against the ‘bounded rationality’ view. We provide qualitative extracts of participants’ reasoning, demonstrating sophisticated and nuanced inferences from detainment to hidden information when the decision is not explained.

How does knowledge of detainment affect juror reasoning?

The intentional binding (IB) phenomenon reflects a perceived attraction between a voluntary action and an intended consequence. Recent research has shown that during IB, intended outcomes expand in time. Although, this effect was significant only for shorter action-outcome delays. However, literature on IB suggests it can also exist for long delays. To address this, we implemented a temporal reproduction task to observe the expansion of an intended outcome. Results revealed the expansion of an intended outcome in the shorter and longer action-outcome delays. These results were discussed under the PIDI (proximal intent distal intent) framework since, using this methodology, both types of intention can be operationalized. Proximal for shorter delays under the motor action, and distal for longer delays as the inference made towards the objective duration to be reproduced. Our implementation allows a robust way to observe the temporal dynamics of an intended outcome, irrespective of the action-outcome delay.

Intended Outcomes Expand In Time: Evidence from the Temporal Reproduction Task

Recent work highlighted a cause’s scope, the number of distinct effects it generates, as important for changing perceived causal strength. The findings, however, have been inconsistent, with researchers offering effect valence and domain knowledge as mutually exclusive explanations for the presence of scope effects. Our pre-registered experiment (N=322) simultaneously controlled or manipulated the links between scope and causal strength to adjudicate between the two accounts. We also examined another structural feature as a potential moderator of a cause’s perceived strength: while prior work has focused on causal power in common cause networks, we compared them with causal chains, where mechanistic clarity may lead to higher causal power ratings. We found that causal strength is robustly higher in chains regardless of valence and domain (p=.009). However, we failed to replicate the prior findings on the scope effect. Our experiment casts doubt on previous authors’ interpretation of scope's impact on causal judgments.

Perceived Causal Strength in Chains vs. Common Causes

Information about an individual’s moral character can bias legal judgments even when it is incidental. The present study explored how evaluations of a defendant’s character can influence mock jurors’ judgments of blame, guilt, and punishment. Participants (N=143) were shown a trial transcript in which the defendant’s character was manipulated by introducing an irrelevant moral behavior observed before the alleged crime as well as prior conviction. Bad defendants were judged more harshly than good or neutral defendants, especially for punishment judgments, where differences were persistent. While character similarly influenced judgments of guilt, blame, and intentionality, the effect diminished as participants received more information about the case. Thematic analysis of judgment rationales also revealed four key themes in participants’ reasoning process: facts of the case, inference of the defendant’s mental states, counterfactual reasoning, and evaluation of the defendant. We discuss the implications of this study in moral and legal decision-making.

Defendant character influences mock juror judgments of blame, guilt, and punishment

This work attempts to bridge the divide between accounts of causal reasoning with respect to agents and objects. We begin by examining the influence of animacy. In a collision-based context, we vary the animacy status of an object using 3D animations. By holding the fine-grained kinematics of the actual and counterfactual outcomes fixed across animate and inanimate conditions, we find that animacy, itself, has no effect on causal attribution judgments. Next, we test if causal judgments for animate and inanimate objects differ as a function of the counterfactuals they respectively afford in a disjunctive causal structure. Here, we find that the effect of perceived animacy on causal attribution is mediated by differences in counterfactual judgments. Finally, we introduce the known effect of prescriptive norm violations to this paradigm. Our findings collectively highlight how normative expectations specify the counterfactual considerations that guide causal reasoning for both agents and objects.

Causal Reasoning Across Agents and Objects

Humans excel at causal reasoning, yet at the same time consistently fail to respect its basic axioms. They seemingly fail to recognize, for instance, that only the direct causes of an event can affect its probability (the Markov condition). How can one explain this paradox? Here we argue that standard normative analyses of causal reasoning mostly apply to the idealized case where the reasoner has perfect confidence in her knowledge of the underlying causal model. Given uncertainty about the correct representation of a causal system, it is not always rational for a reasoner to respect the Markov condition and other ‘normative’ principles. To test whether uncertainty can account for the apparent fallibility of human judgments, we formulate a simple computational model of a rational-but-uncertain causal reasoner. In a re-analysis of a recent causal reasoning study, the model fits the data significantly better than its standard normative counterpart.

Uncertainty can explain apparent mistakes in causal reasoning

Gathering information via question asking is an essential and effective tool for learning. However, it also requires learners to select from a near infinite space of possible queries. Here, we investigate a potentially powerful guide for question asking in young learners: the relationship between cause and effect. Children (5- and 7-year-olds) read a storybook about an event with an unknown cause and made several choices between two questions to ask about possible candidate causes. Both questions revealed similar information, but only one had the potential to determine whether a candidate was capable of causing the event described. Participants overwhelmingly selected causally relevant over irrelevant questions, with strong performance in both age-groups and for all types of information. These results suggest that young learners employ their understanding of the casual connections between events to identify and impose useful constraints on information search.

Children Use Causality to Guide Question Asking

Causal perception is a cornerstone of early cognitive development. A large database of research attests to the fact that infant causal perception emerges between 6 and 10 months of age. However, this research has tended to use artificial stimuli, and it remains unknown how infants learn about the causal properties of more realistic categories such as people and objects. For example, how do infants learn that people can cause other people to act and behave at a distance, whereas inanimate objects require contact to move and act? One answer to this question is that this knowledge is present from birth or shortly thereafter and is underpinned by core knowledge systems. An alternative perspective maintains that infants acquire this knowledge via domain-general associative learning. The goal of the present paper is to demonstrate that this alternative perspective—implemented in a connectionist computational model—is sufficient to explain infants’ developing knowledge about people and object causal action.

How infants learn about people and object causal action: An associative account

Group disparities, such as the gender wage or racial achievement gap, pervade our society. Unfortunately, from a young age, we often attribute these disparities to inherent features of groups, leading to problematic outcomes (e.g. overgeneralizing, justifying group disparities as the norm). Structural thinking, which attributes disparities to a stable external structure, could be an alternative causal framework, although little work has explored its outcomes. Could structural thinking lead to more positive social outcomes than inherent thinking? Adults who learned about the structural (versus biological or cultural) origins of an occupational disparity were more context-sensitive in their generalizations, less likely to endorse the disparity, and provided more structural interventions. 7- to 9-year-olds showed hints of the former two as well. Cultural thinking showed an intermediate pattern between biological and structural thinking. Overall, structural thinking could be a more socially fruitful way of thinking that mobilizes progress on social disparities.

Premium content

Prediction, Explanation, and Control Under Free Exploration

Next from CogSci 2023

Swipe and hold: composing interventions in continuous time causal learning

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES