Abstract: Behavioral evidence suggests that beliefs about causal structure constrain associative learning, determining which stimuli can enter into association, as well as the functional form of that association. Bayesian learning theory provides one mechanism by which structural beliefs can be acquired from experience, but the neural basis of this mechanism is poorly understood. We studied this question with a combination of behavioral, computational, and neuroimaging techniques. Male and female human subjects learned to predict an outcome based on cue and context stimuli while being scanned using fMRI. Using a model-based analysis of the fMRI data, we show that structure learning signals are encoded in posterior parietal cortex, lateral prefrontal cortex, and the frontal pole. These structure learning signals are distinct from associative learning signals. Moreover, representational similarity analysis and information mapping revealed that the multivariate patterns of activity in posterior parietal cortex and anterior insula encode the full posterior distribution over causal structures. Variability in the encoding of the posterior across subjects predicted variability in their subsequent behavioral performance. These results provide evidence for a neural architecture in which structure learning guides the formation of associations.
Introduction
Classical learning theories primarily focus on the "strength" of associations—for instance, how strongly a specific food is linked to a subsequent illness. However, real-world learning is often more complex, requiring the brain to infer the underlying "rules of the game," or causal structure. This process determines how cues, contexts, and outcomes relate to one another. While associative learning (updating links) is well-studied, the neural mechanisms behind structure learning (updating the map) remain less clear. This study by Tomov et al. investigates how the human brain maintains these distinct systems and how they interact to guide behavior.
Experimental Paradigm
To dissociate structure learning from simple association, researchers employed a "health inspector" task. Subjects (n=20 for fMRI) were asked to predict whether a specific food item (the cue) would cause illness based on the restaurant (the context) in which it was served.
The task was structured into nine blocks of training and testing, using custom clipart stimuli. This design allowed researchers to observe how subjects adapted their predictions as the underlying causal rules changed between blocks.
Computational Modeling of Causal Structures
The study utilized a hierarchical Bayesian model to track subject beliefs. This model assumes the brain evaluates three distinct hidden causal structures simultaneously:
- Irrelevant Context (M1): The cue causes the outcome regardless of the context.
- Modulatory Context (M2): The context acts as a switch, determining if the cue leads to the outcome.
- Irrelevant Cue (M3): The context itself causes the outcome, independent of the cue.
By using Kalman filtering to track associative weights and Bayesian updating for the probability of these structures, the researchers found that this structure-learning model fit subject behavior significantly better ($r = 0.97$) than standard Reinforcement Learning models ($r = 0.73$).
Neural Signatures of Structure Learning
Using model-based fMRI, the team identified a distinct frontoparietal network involved in structure learning. They specifically looked for the Kullback–Leibler (KL) divergence, a measure of how much a belief is revised. Updates to causal structure beliefs were uniquely tracked in the posterior parietal cortex (PPC), lateral prefrontal cortex (PFC), and the frontal pole.
Crucially, Representational Similarity Analysis (RSA) revealed that the brain does not simply track the most likely rule. Instead, the PPC and anterior insula represent the full posterior distribution—maintaining the probabilities of all three causal models at once. The "fidelity" or clarity of these neural representations was not just a theoretical finding; it directly predicted how well individual subjects could generalize their knowledge to novel test scenarios.
Conclusion
These findings suggest a neural architecture where structure learning and associative learning are computationally and neurally dissociable. The frontoparietal network acts as a higher-level system that infers the causal rules of an environment, which in turn constrains how associations are formed. This hierarchical organization allows humans to generalize knowledge efficiently across diverse and changing contexts.
Summary updated