Formalizing Anthropomorphism Through Games: A Study in Deep Neural Networks

Martin A. Zinkevich, Dale Schuurmans

Abstract: Anthropomorphization can be used as a tool by humans to reason about complex nonhuman phenomena, by ascribing agents with intelligence and goals that are similar to their own. Deep neural networks are complex structures, and we do not understand well how to optimize them. One way to deal with such complexity is to give complex parts of the network (such as the activation functions) goals and actions, even if these parts are unchanging in their behavior. This allows us to transform the problem of finding the parameters of deep neural networks into a game, and use the same approaches that we use for games to generate good parameters for a deep neural networks. This paper presents the results of (Schuurmans and Zinkevich 2016) to the game theory community.