Why curricula help some learners and not others
Year: 2026
Type: preprint
Abstract: Curriculum learning, the deliberate ordering of learning experiences, is central to how humans and animals learn, yet its benefits in neural networks are inconsistent. We argue that curricula succeed when they prevent or correct early commitments that might otherwise predispose learners to underfit or overfit to training data. Early experience can lead learners to overlook relevant information, or to rely on irrelevant details that do not generalise well. Curricula help by preventing these mistakes or making them easier to correct. Their benefits diminish when learners have spare capacity or suitable representations, and curricula can even harm learning when early experience establishes patterns that later experience cannot build on. The main source of curriculum benefit, thus, is helping learners `unlearn' overly simple or overly complex solutions.
Source: PsyArXiv (OSF Preprints)
Authors: Mohammed Amara, Christopher Summerfield
Institution: University of Oxford
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: green
OpenAlex ID: W7234426302
