[13] Adji Bousso Dieng - Deep Probabilistic Graphical Modeling
The Thesis Review
00:00
How to Minimize the Entropy of Noise in a Generative Model
If we want to prevent the mode collapse, then we should try to maximize the entropy. So I thought it was a really nice solution. How is the entropy computed? Is that expensive to compute? You don't need to compute the entropy. That's a nice thing there too, because we're doing the gradient descent. Ultimately, all you need is to have gradients approximately.Yeah, so you contractively approximate the gradients, because you already are in a situation where you have a sample from the posterior. And so you need to unroll HMC for like to get two samples, two burning samples and then two samples to approximate the expectation.
Transcript
Play full episode
