ConceptGumbel-Softmax / Gumbel Softmax
Gumbel-Softmax
A reparameterization trick that approximates discrete sampling with a continuous relaxation parameterized by temperature, allowing gradients to flow back through discrete selection operations; it is commonly used for optimizing discrete variables.
Timeline (last 90 days)
Oct 22
LoopRec's gating mechanism outputs binary decisions, and during training the non-differentiability problem must be addressed using methods such as Gumbel-Softmax, straight-through estimators, or relaxed sampling with temperature annealing.
Unverified50%