L18.3: Modifying the GAN Loss Function for Practical Use

Sebastian's books: sebastianrasch...
Slides: sebastianrasch...
-------
This video is part of my Introduction of Deep Learning course.
Next video: • L18.4: A GAN for Gener...
The complete playlist: • Intro to Deep Learning...
A handy overview page with links to the materials: sebastianrasch...
-------
If you want to be notified about future videos, please consider subscribing to my channel: / sebastianraschka

[@prateekpatel6082]

Quite vague poor explanation of gradient vanishing when strong discriminator. The gradient -1 is not weak , the reason it vanishes is because -1 is multiplied (chain rule) to gradient of D w.r.t G, which is too small or confident/saturated discrimators. Once you change the formulation to 1/y_hat , the small gradient is not multiplied to a very large number

[@jonathansum9084]

Hi. I have a question. For the second equation at 7:30, how did you get the info for the [0, inf]? Log(0) = -inf. Or is that the equation was changed to be negative log likely hood, so it became inf. Thank you.

[@SebastianRaschka]

Good catch, looks like I forgot the minus sign at the bottom.

[@adityarajora7219]

4:08 how D(G(z)) --> 0 maximize the gradient descent equation??????? please help im completely confused!

[@adityarajora7219]

is minimize meant only magnitude? ....I am considering you are minimizing magnitude! is that the case?

[@SebastianRaschka]

"how D(G(z)) --> 0 maximize the gradient descent equation?" Actually we are not trying to maximize it but minimize it. So in the equation at the top, it's gradient descent with log(1-D(G(z))). Since we feed fake images, D(G(z)) will be close to 0 in the beginning, so the loss is log(1-0) = log(1) = 0. With gradient descent it will go towards -infty as D(G(z)) -> 1.

[@adityarajora7219]

in the end everything is clear.....i have gone through many videos but dint get proper understandig....but this video explains whats going inside those equation!! great

[@SebastianRaschka]

awesome! glad to hear!

[@techgurlpk]

Thank you so much for this beautiful and detailed discussion.. after reading and getting confused from so many places.. landed on your channel .. and it cleared all ambiguities.. 🙏😇

[@SebastianRaschka]

awesome, I am happy to hear this!

[@tiana717]

Thank you for the great video! At 14:19, I understand that the first equation is the gradient ascent from the original paper, and the second equation is negative log-likelihood, which from what I understand is just adding negative signs to the normal log-likelihood. But how did you transform eqn 1 to the normal log-likelihood to begin with? Integration? Thank you!

[@BarryAllen-lh6jg]

well explained. thanks

[@Belishop]

This video is a bit complicated.

[@SebastianRaschka]

Thanks for the feedback!

[@Belishop]

@@SebastianRaschka Thank you for the video!