CausaL

The ability to infer cause-effect relations is an important facet of learning. In particular, learning the causal effect of our behaviors provides the basis for rational decision-making and allows people to engage in meaningful life and social interactions. A number of psychological theories have been proposed to account for the internal representations mediating causal learning. However, an integrated understanding linking causal learning theories and brain systems is still lacking. The CausaL project will investigate the neural and computational bases of causal learning in the context of goal-directed behaviors (action-outcome causal relations). To do so, we will need to lift two key barriers. The first is the lack of neurocomputational models that formalise psychological theories into predictions about the brain computations. The second is the lack of clear understanding of the brain dynamics supporting action-outcome causal learning. Two preliminary studies of our group have set the ground for lifting these barriers and demonstrated the feasibility of the project. The CausaL project will pursue this work along two Tasks. In Task 1, we will confront two leading neurocomputational learning frameworks: Active Inference (AI) and Reinforcement Learning (RL). Both approaches can be used to formalise the relation between decision variables predicted by causal learning theories and learning behaviors in humans, but differ in the underlying conceptual structure. Put simply, whereas the first is a Bayesian approach aiming at the minimization of uncertainty during explorative and exploitative behavior, the latter formalizes learning as a process maximizing cumulative rewards. By means of simulations of ideal agents and comparison with behavioral data collected from a large cohort of participants (N=180), we will develop neurocomputational models of action-outcome causal learning. We will then compare AI and RL models in their ability to accurately explain empirical behavioral patterns, choice patterns and causal scores. Finally, we will exploit the “best” models to yield predictions about the underlying neural computations. In Task 2, we will test AI and RL models on brain data and study how learning-related brain regions interact. In fact, it is now acquainted that action-outcome causal learning is a brain network phenomenon. However, it is still unclear how fronto-striatal regions dynamically interact to support learning computations. We will investigate brain data to test whether the internal representations predicted by psychological theories and implemented in AI and RL models (for example, conditional probabilities between actions and outcomes, causal beliefs) are encoded in: i) functional connectivity dynamics and directional influences between learning-related brain regions, by means of magnetoencephalography studies in healthy participants and intracranial stereo-electroencephalography in epileptic patients; ii) distinct spatial patterns of brain activaty along fronto-striatal territories, by means of functional and diffusion magnetic resonance imaging (MRI). The combination of functional and structural brain data will reveal how causal learning emerges from interplay between large-scale functional interactions and anatomo-functional gradients along fronto-striatal circuits. To conclude, the CausaL project offers the extraordinary opportunity to link theoretical models of action-outcome causal learning, behavior and brain network dynamics, the so-called cognitive architectures of causal learning (CausaL).