A continuous-space recipe for discrete language generation
A post says a growing body of work uses continuous flow models for discrete problems, including language generation.
TLDR
The post outlines three steps: represent vocabulary items as one-hot vectors—one entry is 1 and the rest are 0—learn a flow in that continuous space, then decode at the simplex corners.
Combined views
13.2K
2 Sources, first seen 7h ago
A continuous-space recipe for discrete language generation
A post says a growing body of work uses continuous flow models for discrete problems, including language generation.