Definition
A clever way of training fixed-point models that avoids running the full implicit-gradient computation.
An approximation technique for training implicit models that estimates gradients through the equilibrium without the full inverse-Jacobian solve, used in attractor-model reasoning experiments.