Should Flipout use Rademacher or Uniform Distribution?

[a76yyyy]

In this paper, The original content is "r and s are random vectors whose entries are sampled uniformly from ±1"

I think this should refer to uniform distribution rather than Rademacher distribution.

Is it my understanding errors, or which article explains the effectiveness of the Rademacher Distribution?

[a76yyyy]

Who can answer my questions?

[SiegeLordEx]

The flipout paper uses the rademacher distribution, that is core to the derivation of the algorithm. E.g. in Observation 1:

Let E be a random sign matrix that is independent of ∆dW. Then ∆W = ∆dW ◦ E is identically distributed to ∆dW.

This would not work with a uniform distribution.