How the perception of pain emerges from brain activity remains elusive. Pain perception and associated brain responses are known to be highly sensitive to several contextual, cognitive and mood factors. In particular, prior expectations largely influence pain, as exemplified by powerful placebo and nocebo effects. These effects suggest that pain perception does not result from a direct readout of sensory inputs, but can be seen as the outcome of an inference process based on noisy observations, which necessitates an internal model accounting for endogenous modulations. Yet, there is no unifying framework explaining how prior expectations and incoming stimuli are combined in our brain to lead to pain perception. In this setting, this project aims to shed light on the mechanisms governing the construction and updating of internal pain models. To reach this goal, we will first develop statistical models that can explain how human subjects combine prior information with incoming evidences to identify which inferential features are encoded in brain responses. Second, in order to understand how neuronal circuits can encode these statistical features, we will design artificial neural networks that replicate human performance in pain inference tasks, and analyze how their learned structures help predicting future stimulus intensities. Both families of models will be adapted to the studied context, and assessed on several already recorded electrophysiological datasets. By analyzing the learned reasonings in machine learning models, we aim, on the one hand, to study their artificial learning mechanisms and, on the other hand, to enhance our understanding of biological pain networks.