Estimate the gradient from a small batch instead of the whole dataset. The estimate is noisy, and that noise turns out to help escape sharp minima as well as making each step affordable.
supersedescorrects · extends
classifiesspecializes · part-of
substitutes forapproximates · alternative-to
depends onrequires · validates
Colour is the family; a dashed line is the second member of it.
Drag to pan · scroll to zoom · click a node to open it
This node
correctsfixes a defect in Gradient Descenta full-batch gradient costs the entire dataset per step
Referenced by
correctsBatch Normalization fixes a defect in thiseach layer's input distribution shifts as earlier layers update
requires3D Gaussian Splatting does not work without thisthe primitives are fitted to photographs by gradient descent
correctsMixed-Precision Training fixes a defect in thisfull-precision training is bandwidth bound
correctsMomentum fixes a defect in thisplain SGD oscillates across ravines and crawls along them
requiresNeural Radiance Fields does not work without thisthe field is fitted to photographs by gradient descent
requiresNeural Radiance Cache does not work without thisthe cache is trained online, while rendering