Learning to Attend -- From Bottom-Up to Top-Down

The control of overt visual attention relies on an interplay of bottom-up and top-down mechanisms. Purely bottom-up models may provide a reasonable account of the looking behaviors of young infants, but they cannot accurately account for attention orienting of adults in many natural behaviors. But how do humans learn to incorporate top-down mechanisms into their control of attention? The phenomenon of gaze following, i.e. the ability to infer where someone else is looking and to orient to the same location, offers an interesting window into this question. We review findings on the emergence of gaze following in human infants and present a computational model of the underlying learning processes. The model exhibits a gradual incorporation of top-down cues in the infant's attention control. It explains this process in terms of generic reinforcement learning mechanisms. We conclude that reinforcement learning may be a major driving force behind the incorporation of top-down cues into the control of visual attention.