Fear and anxiety influences on probabilistic learning: A pilot online study and computational modeling

People learn differently under fearful and anxious conditions, especially when the environment itself is volatile. We conducted an online pilot study (as part of the functional magnetic resonance imaging project) on how emotions influence probabilistic associative learning in a volatile environment. Participants had to learn the association between two cues (colour patches) and two outcomes (images of faces or hands), while the reversals occurred at times unknown to participants. Their anxiety state and trait were measured using the ‘State-Trait Anxiety Inventory’ at the end of the experiment. Performance was compared between three different conditions: low, medium, and high fear (using scarified images). Three computational models (Rescorla-Wagner, Hidden Markov Model, and Volatile Kalman Filter) were fitted to the behavioural data in each condition. Bayesian Information Criterion, and cross-validation R-squared were used for model comparison. Our preliminary results suggest that fear-inducing stimuli did not influence learning performance in a volatile environment. However, we speculate that there might be individual differences in participants’ responding to our fear manipulation (bimodal distribution in high fear condition’s performance). As for anxiety, there was a positive correlation between performance and low to moderate STAI total scores (≤100) but not high STAI total scores (≥120), suggesting that moderate amount of anxiety could facilitate learning performance. Finally, model comparison result suggests that a simple model with fixed learning rate (Rescorla-Wagner Model), rather than a complex one with adaptive learning rate (Volatile Kalman Filter) most frequently explained behavioural data. In addition, task structure-related information was rarely utilized in solving our task (Hidden Markov Model)