Prince

icon indicating a website link https://kingofspace0wzz.github.io

icon company Carnegie Mellon University icon location California, Santa Barbara icon location Doing AI research, including but not limited to, machine learning, statistical learning theory, variational inference

Results 3 comments of Prince

Why don't we use MSE as a reconstruction loss for VAE ?

According to the original VAE paper[1], BCE is used because the decoder is implemented by MLP+Sigmoid which can be viewed as a 'Bernoulli distribution'. You can use MSE if you...

Why don't we use MSE as a reconstruction loss for VAE ?

Yes, that is the idea. However, I don't think tanh would be an appropriate choice for the output layer in this case, as we are fitting values in [0,1] rather...

Ensuring the planar flow transformation is invertible

I think it would not be a proper probability distribution if the flow is not invertible, as you need the determinant of inverse jacobian to transform a random variable.