At the heart of statistical learning lies the concept of uncertainty.Similarly, embodied agents such as robotsand animals must likewise address uncertainty, as sensationis always only a partial reflection of reality. Thisthesis addresses the role that uncertainty can play ina central building block of intelligence: categorization.Cognitive agents are able to perform tasks like categorical perceptionthrough physical interaction (active categorical perception; ACP),or passively at a distance (distal categorical perception; DCP).It is possible that the former scaffolds the learning ofthe latter. However, it is unclear whether DCP indeed scaffoldsACP in humans and animals, nor how a robot could be trainedto likewise learn DCP from ACP. Here we demonstrate a methodfor doing so which involves uncertainty: robots performACP when uncertain and DCP when certain.Furthermore, we demonstrate that robots trainedin such a manner are more competent at categorizing novelobjects than robots trained to categorize in other ways.This suggests that such a mechanism would also beuseful for humans and animals, suggesting that theymay be employing some version of this mechanism.
展开▼