Deep Generative Modeling
The latest 2021 lectures from the MIT Introduction to Deep Learning class are trickling out into the known universe. And we have a great one to watch today. Ava Soleimary will school us on all things generative. Well not all things, but it's pretty good overview.
She covers VAE architecture, and i was struck by her description of how it works and LeCun's Energy Model unified field theory of generative models. Its fascinating to compare the 2 descriptions.
And regularization is all about centering the mean while regularizing the variance. And when you think about it that way it seems a lot more straightforward than you might have thought at first.
Then we get into the latent space, latent perturbation and disentanglement, GANs, a really great intuition description of how GANs are transforming one distribution into another distribution, StyleGan, CycleGAN, etc.