Unified Neural Scaling Laws

Researchers propose a Unified Neural Scaling Law to predict deep neural network performance when model size, data, training steps, compute, and hyperparameters change together.

The method shows better extrapolation accuracy than previous scaling formulas across vision, language, math, and reinforcement learning tasks.

A more reliable scaling law could help teams plan experiments, compare AI systems, and estimate returns from more data or compute.

The work is shared as an arXiv paper and targets performance prediction across modern deep learning setups.