Search Papers | Poster Sessions | All Posters
Poster A95 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Human Curriculum Effects Emerge with In-Context Learning in Neural Networks
Jacob Russin1,2 (), Ellie Pavlick1, Michael J. Frank2,3; 1Department of Computer Science, Brown University, 2Department of Cognitive and Psychological Sciences, Brown University, 3Carney Institute for Brain Science
In tasks governed by succinct rules, human learning is more robust when related examples are blocked, but in the absence of such rules, interleaving is more effective. To date, no neural model has simultaneously captured these seemingly contradictory effects. Here we show that these effects spontaneously emerge in neural networks capable of “in-context learning” (ICL). In both language models and metalearning networks, ICL explains the observed blocking advantage while concurrent in-weight learning explains the interleaving advantage.
Keywords: neural networks language models metalearning curriculum