Can reinforcement learning explain the development of causal inference in multisensory integration?

Bayesian inference techniques have been used to understand the performance of human subjects on a large number of sensory tasks. Particularly, it has been shown that humans integrate sensory inputs from multiple cues in an optimal way in many conditions. Recently it has also been proposed that causal inference can well describe the way humans select the most plausible model for a given input. It is still unclear how those problems are solved in the brain. Also, considering that infants do not yet behave as ideal observers, it is interesting to ask how the related abilities can develop. We present a reinforcement learning approach to this problem. An orienting task is used in which we reward the model for a correct movement to the origin of noisy audio visual signals. We show that the model learns to do cue-integration and model selection, in this case inferring the number of objects. Its behaviour also includes differences in reliability between the two modalities. All of that comes without any prior knowledge by simple interaction with the environment.