Memory-Augmented Theory of Mind Network

Social reasoning necessitates the capacity of theory mind (ToM), ability to contextualise and attribute mental states others without having access their internal cognitive structure. Recent machine learning approaches ToM have demonstrated that we can train observer read past present behaviours other agents infer beliefs (including false about things no longer exist), goals, intentions future actions. The challenges arise when behavioural space is complex, demanding skilful navigation for rapidly changing contexts an extended period. We tackle by equipping with novel neural memory mechanisms encode, hierarchical attention selectively retrieve information others. memories allow rapid, selective querying distal related deliberatively reason current state, behaviours. This results in ToMMY, a model learns while making little assumptions underlying processes. also construct new suite experiments demonstrate facilitate process achieve better performance, especially high-demand false-belief tasks require inferring through multiple steps changes.