Skip to content
Synced · WeChat

Mila and DeepMind Propose UNSL for Unified Multivariate Neural Scaling Laws

多变量神经缩放定律迈向大一统:Mila联手DeepMind提出UNSL

Mila and Google DeepMind proposed Unified Neural Scaling Law, a multivariate scaling-law form that models parameter count, token count, training steps, bottlenecks, overfitting, and adverse hyperparameter effects; UNSL achieved the best extrapolation on 60.87% of vision tasks and 88.89% of language tasks in the reported experiments.

Why it matters: HKR-H/K/R all pass: UNSL unifies parameters, tokens, steps, bottlenecks, overfitting, and hyperparameter feedback, with vision/language extrapolation numbers. Technical density keeps it in the 78–84 band.

Read the original ↗Export Markdown