Papers of the day   All papers

Gradient Estimation with Stochastic Softmax Tricks


Chris J. Maddison: Stochastic Softmax Tricks: We generalize the Gumbel-Softmax, and introduce new tricks for backpropping through all kinds of random discrete objects: spanning trees, subset selection, & more ( First authors @mbpaulus, @damichoi95.

5 replies, 1055 likes

will grathwohl: HOT SHIT ALERT

2 replies, 44 likes

Thomas Kipf: With Stochastic Softmax Tricks (, Neural Relational Inference can be used to discover other forms of latent structure, such as spanning trees. Figure from the Stochastic Softmax Tricks paper.

0 replies, 35 likes

Dr Simon Osindero: Very cool! Further expanding the set of backprop-able building blocks for model architectures.

1 replies, 9 likes

HotComputerScience: Most popular computer science paper of the day: "Gradient Estimation with Stochastic Softmax Tricks"

0 replies, 2 likes

arXiv in review: #NeurIPS2020 Gradient Estimation with Stochastic Softmax Tricks. (arXiv:2006.08063v1 [stat\.ML])

0 replies, 1 likes


Found on Jun 16 2020 at

PDF content of a computer science paper: Gradient Estimation with Stochastic Softmax Tricks