About Reparameterization Trick of spike variable "gamma" about variational-sparse-coding

[Alfo5123]

Hi @Ronosu thanks for your comment.

Indeed, formula 12 of the paper (https://openreview.net/pdf?id=SkeJ6iR9Km) shows what you mentioned. As you can see in the code from 'vsc.py' we write, '-F.relu(-self.fc23(h1))'. That is, we pass the output through a fully connected layer before changing the sign and passing through activation. This causes the value to not be zero, and carry information.

Regarding the second question, it was not our decision, it was suggested by the authors of the paper as you can see in the section C2 of the appendix, in the link above.