Saurab Gyawali

DALE investigates whether the way children learn can make machine learning more efficient. Using a contrastive learning task in which participants pair spoken Danish animal words with images, we quantify how the benefit of adaptive versus random learning changes from childhood through adulthood. The order and pace at which learners master each word are then used to build epoch-specific training curricula for vision-language models: we fine-tune CLIP under adaptive, random, and mixed curricula and compare them against standard training regimes, measuring both task performance and how closely the model's internal representations come to resemble human ones. The broader question is whether human developmental learning trajectories carry structure that is genuinely useful to artificial systems, and whether AI can serve as a testable model of human cognitive development.