MIT 6.S191 (2021): Deep Generative Modeling

MIT 6.S191 (2021): Introduction to Deep Learning
Deep Generative Modeling
Lecturer: Ava Soleimany
January 2021

Lecture Outline
0:00​ - Introduction
6:03 - Why care about generative models?
8:56​ - Latent variable models
11:31​ - Autoencoders
17:00​ - Variational autoencoders
24:30 - Priors on the latent distribution
34:38​ - Reparameterization trick
38:14​ - Latent perturbation and disentanglement
41:25 - Debiasing with VAEs
43:42​ - Generative adversarial networks
46:14​ - Intuitions behind GANs
48:27 - Training GANs
52:57 - GANs: Recent advances
57:15 - CycleGAN of unpaired translation
1:01:01​ - Summary

Subscribe to stay up to date with new deep learning lectures at MIT, or follow us @MITDeepLearning on Twitter and Instagram to stay fully-connected!!

[samiswilf]

She uses Plato's Allegory of the Cave to explain latent spaces. I like it!

[GovindJeevan]

These videos are a great service to humanity, truly! Democratizing all this quality Deep Learning content enables so, so many of us!

[joshcummins]

My goodness. You teach so clearly. I appreciate the slides as well. Been working on Gans and autoencoders to make art. This really summarized everything in an easy way to understand.

[ajaytaneja]

Love the first slide - which face is real! Inspires you watch this lecture "n" times until you understand
Thank you Alexander and Ama

[ariG]

Great lecture! Thanks for the intuition of the prior and the regularisation term used in VAE.

[RamanVenu]

Another well done video! The quality of these lectures make me wish I actually attended MIT

[jpark]

49:52
Am I the only person confused about this?
I guess the equation should be like E_x[log D(x)] + E_z[log(1-D(G(z)))]?

The equation in the video is E_z[log D(G(z))] + E_x[log(1-D(x))] and if we are trying to maximize this equation,
D(G(z)) should be 1 and D(x) should be 0 which means that the discriminator says "the fake data is real" and "the real data is fake"
The discriminator must say "the fake data is fake" and "the real data is real" to maximize the equation.

If the equation is E_x[log D(x)] + E_z[log(1-D(G(z)))] (my suggestion), the equation will be optimized by maximizing the equation because
log D(x) = 0 and log (1-D(G(z))) = 0 which means that D(x) = 1 and D(G(z)) = 0. Finally this could be interpreted as "real data is real" and "fake data is fake" which suites our original goal.

Can anyone help me about this equation? Thanks

[reandov]

Another great class! Generative Adversarial Networks are by far, for me at least, the most fun type of Neural Networks and the sparse number of applications allowed by them are huge! Huge thanks to Ava Soleimany for the class, as well to everyone in the project! It's my first time so far in this introductory course and I'm loving it!

[balaganesh]

If you can share the details reference text books or PDFs or any blogs to go little bit deeper or to understand the math behind it, will be helpful to us. BTW these series are awesome. Thank You for sharing. Always ❤️ MIT

[kqcdbdo]

Already 44 likes before beginning haha. This shows how popular your videos are!

[PedramNG]

WoW, I enjoyed every second of this lecture! 😃

[bender]

Great video! Actually there's different types of deep generactive model proposed recently besides vae and gan. In terms of density estimation, normalizing flows might be a better solution than vae. Score based generative models are recently proposed and have great performance on image generation.

[chaoukimachreki]

I understood everything prior to plato's parable.

[un_feffo]

This debiasing procedure seems so useful that it could be applied to every classification problem.. Or not?

[mr_knowitall]

What software are you using for the spectograms? Where can we learn more about voice synthesis?

[jianhuajin]

Help! I copy the URL to the colab, but I can't find the lab2 and lab3.only lab1 part1 and part2. Is there anything wrong?

[samirandas]

This series is simply the best. So clear and informative. I wonder are there measures to identify/ measure the biasness in a DL model? It is an interesting line of research,

[mickmickymick]

In my dream world, I would be good enough to go out with Ava Soleimany

[lackushi]

To the 5 people who disliked this video.


What!?

[SuperHoggs]

Thank you so much! ps. did anyone notice how the background is a green screen :D