EasyDeepLearn

What is gradient noise scale (GNS) and how do you use it?

hard

Answer

  • GNS (McCandlish et al., 2018) measures the ratio between the variance of the mini-batch gradient and its squared mean, roughly telling you how much batch size you can grow before returns diminish.
  • Compute Bcrit  =  trace(Cov(g))B_{\mathrm{crit}}\; = \;\mathrm{trace}(\operatorname{Cov}(g)) / ||g||².
  • Below BcritB_{\mathrm{crit}}, doubling batch halves steps.
  • Above, benefits plateau.
  • Guides choice of batch size for large-model training.
Check yourself — multiple choice
  • GNS is a measure of activation variance
  • Ratio Tr(Cov(g)) / ||g||² — tells you the critical batch size beyond which scaling gives diminishing returns
  • GNS is an activation function
  • GNS is a loss

GNS: Tr(Cov(g))/||g||² = critical batch size — guides scaling of large-model batches.

#batch-size#distributed#training

Practise Deep Learning

214 interview questions in this topic.

Related questions