Abstract

This week we will read about “Machine Theory of Mind”, by Rabinowitz et al. who describe a neural network, their ToMnet, which autonomously learns how to model other agents’ behaviour, beliefs and intentions. ToMnet even recognises an important separation in the Theory of Mind literature between the beliefs that other agents hold about the environment and its own.

Which features of their design allow ToMnet to infer the characteristics of other agents so successfully in their Gridworld experiments? And can we think of any challenges TomNet would face in larger multi-agent environments?

Paper: https://arxiv.org/abs/1802.07740

We strongly recommend you read the paper before coming. Regardless, from 17:00 we will be reading the paper and people can arrive. The discussion proper will start at 17:30. There will be pizza!