<ul><li>Scaling has been driving advancements in deep learning with many studies identifying power-law scaling laws.</li><li>Existing scaling law prediction methods lack uncertainty quantification crucial for decision-making.</li><li>A Bayesian framework using Prior-data Fitted Networks (PFNs) is proposed for neural scaling law extrapolation.</li><li>The method shows superior performance in extrapolation and Bayesian active learning scenarios, offering uncertainty-aware predictions.</li></ul>

Bayesian Neural Scaling Laws Extrapolation with Prior-Fitted Networks

Discover more