Human Pavlovian fear conditioning conforms to probabilistic learning

Learning to predict threat from environmental cues is a fundamental skill in changing environments. This aversive learning process is exemplified by Pavlovian threat conditioning. Despite a plethora of studies on the neural mechanisms supporting the formation of associations between neutral and aversive events, our computational understanding of this process is fragmented. Importantly, different computational models give rise to different and partly opposing predictions for the trial-by-trial dynamics of learning, for example expressed in the activity of the autonomic nervous system (ANS). Here, we investigate human ANS responses to conditioned stimuli during Pavlovian fear conditioning. To obtain precise, trial-by-trial, single-subject estimates of ANS responses, we build on a statistical framework for psychophysiological modelling. We then consider previously proposed non-probabilistic models, a simple probabilistic model, and non-learning models, as well as different observation functions to link learning models with ANS activity. Across three experiments, and both for skin conductance (SCR) and pupil size responses (PSR), a probabilistic learning model best explains ANS responses. Notably, SCR and PSR reflect different quantities of the same model: SCR track a mixture of expected outcome and uncertainty, while PSR track expected outcome alone. In summary, by combining psychophysiological modelling with computational learning theory, we provide systematic evidence that the formation and maintenance of Pavlovian threat predictions in humans may rely on probabilistic inference and includes estimation of uncertainty. This could inform theories of neural implementation of aversive learning.Author summary: Using environmental cues to predict threat is a crucial skill, encountered in many species. A laboratory model to study associating predictive cues with aversive events is Pavlovian fear conditioning. A computational understanding of this process at a systems-level is lacking. Here, we investigate which computational learning model best predicts activity of the human autonomic nervous system during fear conditioning. We show, across three data sets and two autonomic readouts, that a probabilistic learning model explains the data decisively better than previously proposed models. We suggest that humans learn to predict threat by maintaining and constantly updating a probabilistic model of the environment through Bayes’ rule.