Learning Fast and Slow: Accelerating AI Training with Human-inspired Learning
Navn på bevillingshaver
Max Müller-Eberstein
Titel
Postdoctoral Fellow
Institution
University of Tokyo
Beløb
DKK 1,673,658
År
2025
Bevillingstype
Internationalisation Fellowships
Hvad?
This project makes AI training more efficient by having them learn more like humans. Instead of slowly training one large model on a mix of everything - which leads to worse performance in languages with less data - it trains smaller models faster on data, which we know humans could learn a skill from. By combining models with complementary skills, we then enable new applications in smaller languages.
Hvorfor?
Compared to humans, current AI systems learn slow: requiring multiple thousand lifetimes worth of data to train. They also struggle to combine different skills, being worse at the same math problem if asked in a 'smaller' language. To uncouple what AI models can do from the language they work in, we need a modular approach that learns as fast as humans do.
Hvordan?
Based on theories of how humans learn specific skills, this project develops efficient training strategies for showing AI models only the most relevant information required to solve each particular task. By comparing what each task-specific model learns, and re-combining the best performing ones, we create new models, which have a higher skill level across a broader set of tasks and languages.