About the Event
How can agents learn to cooperate? Mammalian brains are stunningly successful examples of populations of billions of agents that learn to cooperate on the basis of extremely limited information. In this talk, I will discuss recent work analyzing neuronal learning using methods from learning theory and mechanism design.
Firstly, I show that discretizing standard (textbook!) models of synaptic plasticity leads to a model of neurons as rational agents that maximize biologically plausible objective functions. Secondly, I analyze the incentives encoded in neuronal objective functions, focusing on a striking mathematical connection between neurons and prediction markets. Finally, I sketch work in progress applying these ideas to semi-supervised learning and distributed representations.