Technical Program

Paper Detail

Paper: PS-1A.32
Session: Poster Session 1A
Location: Symphony/Overture
Session Time: Thursday, September 6, 16:30 - 18:30
Presentation Time:Thursday, September 6, 16:30 - 18:30
Presentation: Poster
Paper Title: Generalized Schema Learning by Neural Networks
Manuscript:  Click here to view manuscript
Authors: Catherine Chen, Qihong Lu, Andre Beukers, Chris Baldassano, Kenneth Norman, Princeton University, United States
Abstract: Humans have schematic knowledge of how certain types of events unfold (e.g., restaurant meals) that can readily be generalized to new instances of those events. Here, we test whether neural networks can do this kind of generalized schema learning. We stochastically generate stories according to predefined rules and test networks’ ability to answer questions about role fillers (e.g., who is the waiter). We find that networks trained on a small set of fillers can only generalize to stories using the same fillers. By training on a large number of fillers represented by random vectors, we allow networks to generalize to fillers they have never seen before. We find qualitative differences in learning ability between networks with different forms of memory, with some networks learning fewer categories of tasks than others. We further find distinct influences of task difficulty on learning order, and of training order on learning ability.