There is increasing evidence suggesting, that people encode dynamic visual information probabilistically. However, the mechanism of this phenomena is unknown. Recently Kidd et al (2012) investigated, how predictability of varying visual stimuli influences attention and learning in infants. Their main finding was, that infants maintain attention longest for stimuli, that have intermediate predictability. Such a behavior could be explained by applying an optimal learning mechanism, where attention is allocated at the most informative stimuli. However, this prediction could not be explicitly tested by Kidd et al as their method did not have a separate measure of attention and learning. To explore this prediction, we investigated how people perceive and learn about probabilistic events, and tested how accurately their behavior could be captured by a probabilistic framework. In our set of experiments we measured how precisely people can estimate the probabilities of multiple, intertwined simple visual events. We found that increased variability, due to multiple shapes can lead to better estimation. This corroborates previous findings, that probabilistic processes are better captured by implicit than by explicit mechanisms. We also found a linear relationship between visual probabilities and participants’ estimates. Moreover, the pattern of learning within and across blocks is well predicted by a probabilistic model, that includes the statistical structure of the input of the task. These results support the notion, that human learning of dynamic events is well captured by a framework that assumes probabilistic encoding.