How do humans infer the causes and consequences of their actions in complex, uncertain environments? Effective decision-making depends on the ability to link outcomes to the actions or events that caused them. While this has been coined as causal inference in cognitive science, the field of machine learning refers to this ability as credit assignment.

This symposium combines computational modeling, behavioral experiments, neuroimaging, and patient studies to explore the mechanisms of credit assignment. First, Franz Wurm will discuss the importance of prediction errors and dopaminergic signaling and how a tonic manipulation via the dopaminergic precursor tyrosine affects successful credit assignment and surprise minimization. Jutta Peterburs will talk about how cerebellar dysfunction leads to impaired prediction error signaling and suggests weakened credit assignment and cognitive control during feedback and error processing. The credit assignment problem becomes especially acute in social situations with multiple agents and conflicting goals. Marco Wittmann will present insights on how abstract representations form fundamental building blocks for social-decision making and allow quick and accurate causal inference. André Forster presents EEG and computational modeling data exploring how individuals learn to distinguish between task difficulty and their capabilities in a competitive social situation. Lisa Spiering will discuss how people use behavioral flexibility to better assign outcomes to themselves versus others during cooperation and how these processes go awry in depression.

In conclusion, this symposium aims to bring together researchers studying decision-making, learning, social cognition, and those working at the intersection of machine learning, neuroscience, and clinical applications.

The credit assignment problem arises when the link between causes and consequences is not directly observable. In a previous study we highlighted the central role of prediction errors (i.e., surprise) to establish a correct representation of environmental contingencies and thereby solve the credit assignment problem. As prediction errors have been closely linked to catecholaminergic responses, the current study asks if catecholaminergic modulation is not only involved in value learning but also drives structure learning and causal inference. Therefore, we administered a single dose of the catecholamine precursor tyrosine on our novel variant of the bandit task, using a double-blind, placebo-controlled between-subjects design (n = 48). We replicate basic behavioral, computational and neural findings, showing the robustness and reliability of the task design and computational approach. However, contrary to our expectations, first analysis suggest that tyrosine does not alter markers of credit assignment. On the behavioral level, neither hallmarks of implicit nor explicit credit assignment did differ between treatment groups. On the computational level, we find converging evidence for tyrosine’s role in choice stochasticity/greediness but no indication of its impact on structure learning. While single-trial latent variable EEG analysis confirms the neural patterns of credit assignment, there is no evidence for catecholaminergic modulation. Taken together, these preliminary findings suggest that tyrosine impacts certain aspects of decision-making, but it does not appear to specifically influence credit assignment. We will further investigate baseline-dependent effects and discuss the role of tonic and phasic levels for value and structure learning.

Humans are extremely proficient in tracking the value of specific stimuli and adapting their behavior based on their reinforcement history. Herein, prediction errors (PEs), i.e., mismatches between actual and expected action outcomes, are critical. While processing of PEs in reinforcement learning (RL-PEs) has traditionally been linked to cerebral regions such as the striatum and anterior cingulate cortex, recent rodent data suggest that the cerebellum, which is typically associated with processing sensory PEs, also processes RL-PEs. In a series of EEG experiments with a probabilistic feedback learning task, we tested whether cerebellar output is necessary for cerebral RL-PE processing as reflected in the feedback-related negativity (FRN) in the event-related potential. In Experiment 1, 26 patients with chronic cerebellar stroke and 26 matched healthy controls were tested. In Experiment 2, single-pulse cerebellar transcranial magnetic stimulation (TMS) was applied in healthy participants (n=24), thus implementing a virtual lesion approach. No significant RL-PE processing was observed in the FRN in cerebellar stroke patients, and in healthy participants receiving cerebellar TMS. These results show that RL-PE processing in the forebrain depends on cerebellar output. Specifically, cerebellar PE signals appear to directly modulate reinforcement learning, with cerebellar dysfunction disrupting this gating signal, possibly leading to weakened credit assignment. Even though overall learning performance was preserved, reduced behavioral flexibility following cerebellar dysfunction can also be interpreted in the context of weakened credit assignment.

Effective causal inference in social contexts requires tracking one's own actions and distinguishing them from others' actions. I will present a neuro-computational framework examining how the brain solves this credit assignment problem during social interactions, with particular focus on the phenomenon of "self-other mergence" where individuals fail to maintain distinct representations of self and other. Using neuroimaging, brain stimulation, and computational modelling, I will show both correlational and causal evidence that dorsomedial prefrontal cortex plays a critical role in calibrating self-other representations and managing appropriate credit assignment. I will present a broader theory suggesting medial prefrontal cortex accomplishes this by compressing information in social situations, creating abstract representations that serve as building blocks for efficient social decision-making and causal inference. Intriguingly, our model suggests self-other mergence – the apparent misattribution of actions to the inappropriate player - may not reflect a failure of learning mechanisms but rather emerges as an byproduct of this adaptive computational mechanism. The suggested information compression mechanism can enhance overall decision-making accuracy, albeit at the cost of occasionally blurred self-other boundaries when assigning credit for outcomes. This work suggests that abstract neural representations support adaptive decision-making and that credit assignment mechanisms should be contextualized within the broader social dynamics essential for navigating complex social environments with multiple agents and potentially conflicting goals.

A key feature of many psychological disorders is a biased attribution style in which negative outcomes are attributed to internal, uncontrollable traits, while positive outcomes are considered the result of external, similarly uncontrollable factors. A fundamental problem with this maladaptive credit assignment is the distinction between one's own abilities and the demands/difficulty of a given situation. Once individuals develop the (potentially false) belief that they are incompetent, positive outcomes -rather than challenging this belief- may instead be interpreted as evidence that the task was just very easy. This may reinforce maladaptive self-concepts and prevent adaptive belief updating. This talk presents a gambling task designed to investigate this process. Participants competed against computer opponents by placing bets, winning if their offer exceeded that of their opponents. However, much like real-world situations, they could only decide how much effort to exert when placing their bet without knowing the actual amount they were betting. Similarly, their opponents' bets remained hidden. Crucially, participants’ overall ability fluctuated over time, influencing the (hidden) impact of their effort. This design allowed us to examine how individuals adjust their beliefs about their own ability versus their perception of opponent/task difficulty in response to changing win/loss outcomes at different effort levels. The study was conducted in two phases: an online pilot study that informed a computational modeling approach predicting participants' effort and their belief-updating, and an EEG-study of the same task testing electrophysiological correlates of model estimates and providing insight into the neural correlates of learning under uncertainty in credit assignment.

A key human ability is to learn from outcomes by identifying their cause (“assigning credit”), and to figure out whether one’s actions have any influence on our environment. In this study, firstly we examined the cognitive and neural mechanisms of how participants assign credit to themselves versus others and estimate their control over the environment. Secondly, we investigated how these processes can go awry in individuals with depressive symptoms. For this, we developed a novel behavioural paradigm, in which participants inferred from feedback how well they and another player perform, and how much control they exert. We employed this paradigm in a 3T fMRI study (n=50). Participants were recruited based on their depression scores, ranging from no to severe depressive symptoms. We found that to better disambiguate feedback, people engaged in exploratory behaviour (“active disambiguation”, AD). During AD, people changed their performance to test the effect of their behaviour on the feedback, thereby inferring their control. Activity in the supramarginal gyrus at the time of outcome was related to assigning feedback to estimating control from AD and tracking outcomes in an agent-based manner. Preliminary results indicate differences in how individuals assigned outcomes to themselves and others, related to depression symptoms. Next, we will examine the neural correlates of these interindividual differences. These findings may help develop new treatments for depression, such as non-invasive brain stimulation.