Search Papers | Poster Sessions | All Posters
Poster B50 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
Emergent human-like path preferences and implicit subgoal selection in transformers learning graph traversal
Yuxuan Li1, James L. McClelland1; 1Stanford University
Cognitive scientists have proposed normative and heuristic principles that describe human subgoal choices and their partitioning of problems into smaller ones. Here we study the processes through which these choices or partitions arise. Building on the graph-based tasks from prior work, we train neural networks on shortest-path traversal to test whether human-like task decomposition emerges over learning. We find that a simple transformer develops a preference for paths containing nodes that occur frequently on the shortest paths in the graph, consistent with human subgoal preferences. This preference is strongest early in model learning, a phenomenon that might also be observed in human learners. We also find evidence of implicit subgoal selection in the models. These results lay the ground for using neural networks to study how humans learn to decompose tasks and select subgoals by integrating over relevant experiences.
Keywords: cognition learning subgoal discovery neural networks