Reverse-mode differentiation over the network graph: one backward pass yields the gradient with respect to every parameter, at roughly the cost of the forward pass.
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
requiresdoes not work without Gradient Descentit supplies the gradient the update step consumes
Referenced by
requiresCross-Entropy Loss does not work without thisit is the scalar the backward pass differentiates
correctsDropout fixes a defect in thisunits co-adapt and memorise instead of generalising
correctsGradient Checkpointing fixes a defect in thiskeeping every activation exhausts memory
requiresLearned Denoising does not work without thisthe filter is learned from noisy and converged pairs
correctsReLU fixes a defect in thissaturating activations drive the gradient toward zero
correctsResidual Connection fixes a defect in thisgradients vanish through a deep stack of layers