I'm trying to understand and implement the factor-VAE and I'm using the paper Disentangling by Factorising. My probability background is weak and therefore I have issues understanding basic concepts, such as this one:


z is in this case the representation given by the encoder, which in my implementation is 10, that is, z$\in \mathcal{R}^{10,1}$. However, it makes no sense, as far as I can tell, to multiply the encoded representations together to get the $\hat{q}$ in my case, since this would correspond to only a scalar and not the vector that I'm interested in. I'd like to believe that each $z_{j}$ comes from a distribution, which may or may not be correlated to the others, and if I sample from all those fictional distributions, I'd somehow end up with $\hat{q}(z)$. Can someone provide me more information regarding this?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.