[13] Adji Bousso Dieng - Deep Probabilistic Graphical Modeling
The Thesis Review
00:00
The Difference Between Normal View and Probabilistic Graph View of the GAN
The origin of view of the GAN is to say, let's learn to sample. We don't need to be imagining any densities or any model. One way to sample is to take noise, pass it through a neural network and take the output. And learn that with adversarial training. So I have to introduce this auxiliary neural network called a discriminator That will try to put high likelihood likelihood on my observed data and low likelihood, sorry, on my generative samples. When you do that, you're limited, because then you don't have a density over your samples.
Play episode from 52:22
Transcript
