TL;DR
This paper proposes a computational model of causal learning inspired by the "Neurath's ship" metaphor, suggesting that humans learn by making small, local updates to a single, current hypothesis rather than maintaining a full distribution of possibilities. This "Neurath's ship" model, combined with a local uncertainty-based strategy for choosing interventions, effectively explains human behavior in causal learning tasks, offering a more cognitively plausible and scalable alternative to exact Bayesian inference.
What Problem Does This Paper Address?
Learning causal relationships in complex environments is computationally demanding. Exact Bayesian inference, which would ideally identify the true causal structure, becomes intractable as the number of variables increases. This suggests that human causal learning must rely on approximations. The paper addresses how humans learn causal structures and select informative interventions under these computational constraints, drawing inspiration from the "Neurath's ship" metaphor. This metaphor posits that scientific theories (and by extension, our understanding of the world) are continuously rebuilt, with new ideas integrated by modifying existing structures rather than starting from scratch. The challenge is to formalize this intuitive idea into an algorithmic model that explains human behavior.
Methods
The authors propose an algorithmic-level model of causal structure learning based on the "Neurath's ship" metaphor. Key aspects include:
- Single Hypothesis Representation: Learners maintain only a single global hypothesis about the causal structure at any given time, rather than a distribution over all possible models.
- Local Updates: This hypothesis is updated incrementally through "local edits" (adding, removing, or reversing edges) that are conditioned on the rest of the current model. This process is inspired by algorithms like Gibbs sampling and stochastic hill climbing from machine learning.
- Sequential Belief Change: The model formalizes this as a process where learners iteratively refine their current hypothesis based on new evidence.
- Local Uncertainty Schema for Intervention Selection: To address the computational cost of choosing optimal interventions, the paper proposes a two-stage process: first, identifying a "local focus" (e.g., a specific edge or variable's effects) based on current uncertainty, and second, selecting an informative intervention within that focus. This contrasts with globally optimal intervention selection.
Datasets
The paper evaluates its model using data from three experiments:
- Bramley, Lagnado, and Speekenbrink (2015): Participants learned causal structures in 3-variable systems.
- Experiment 1 (new): Explores learning in more complex 3- and 4-variable problems with varying strengths and noise levels.
- Experiment 2 (new): Investigates learning when causal strengths and noise levels are unknown, and also elicits participants' confidence judgments and free explanations for their intervention choices.
No new datasets were introduced; the focus was on analyzing existing and newly collected behavioral data.
Key Insights and Ideas
- "Neurath's Ship" as a Cognitive Algorithm: The paper successfully formalizes the intuitive "Neurath's ship" metaphor into a computational model, suggesting that human causal learning is a process of continuous, local refinement of a single hypothesis.
- Local Focus for Intervention Selection: Humans likely don't aim for globally optimal interventions but rather focus on resolving uncertainty in specific, manageable parts of the causal structure. This "local uncertainty schema" explains why interventions might appear less efficient than ideal but are still informative.
- Sequential Dependence is Key: The model's ability to capture the "stickiness" and recency effects observed in human judgments highlights the importance of sequential updates and the influence of prior beliefs.
- Trade-off Between Accuracy and Cognitive Load: The model provides a framework for understanding how humans balance the need for accurate causal models with their limited cognitive resources, leading to approximate but effective learning strategies.
Key Results
- The "Neurath's ship" model (NS) provided a good fit to participants' judgment patterns across the experiments, often outperforming other heuristic models like "win-stay, lose-sample" (WSLS) and "simple endorsement" (SE).
- In model fitting, NS was the best-fitting model for Bramley, Lagnado, and Speekenbrink (2015) and Experiment 1, while SE was best for Experiment 2. All three heuristics substantially beat an exact Bayesian inference account.
- Participants' intervention choices were more informative than random selection but less so than ideal active learning, and were consistent with the proposed local uncertainty schema.
- Experiment 2 revealed that participants could effectively learn causal structures even without prior knowledge of parameters, and their intervention choices showed a tendency towards more constrained tests when uncertainty was higher.
- Free-response explanations in Experiment 2 suggested that participants focused on specific aspects of the problem (e.g., individual connections or effects) rather than global uncertainty.
Limitations and Future Work
- Measurement Effects: The authors acknowledge that eliciting judgments repeatedly might influence learning processes. They attempted to mitigate this by incentivizing participants and varying reporting conditions, but note that conflating genuine incremental change with response laziness remains a concern.
- Acyclicity Assumption: The model assumes directed acyclic graphs (DAGs) for causal representation, which simplifies inference but may not fully capture real-world phenomena involving cyclic or reciprocal relationships. Future work could explore extensions to dynamic Bayesian networks or chain graphs.
- Noisy-OR Functional Form: The assumption of a noisy-OR functional form for causal relationships might be a simplification. Future work could explore more diverse functional forms or domains where priors are better understood.
- Antifoundationalism: The strong assumption that people consider only a single global hypothesis might be too restrictive for simpler problems where global reasoning might be feasible.
- Selective Memory: The model's assumption of using only recent data is a simplification. Future studies could investigate what specific evidence learners remember and how it's integrated.
- Intervention Choice Modeling: While the local uncertainty schema captures general patterns, modeling the precise choice of focus and intervention remains an area for further development, particularly in understanding how learners estimate local confidence.
Summary updated