5. ML vs Adversarial
• Value function – not cost function
• Min-max
• Player1 wants to mimimise the winning option of player2 , player2
want to maximize the winning option
• Player 1 looking for local minima , player 2 looking for local maxima
• Nash equllibrium
• Each player is assumed to know the equilibrium strategies of the other
players
6.
7.
8. ICLR2019 talk by
Ian GoodFellow
Progressive GAN
2 representative models created from celebA dataset
Take the sample and learn the probability distribution-
generate new samples from the same distribution
• Statistically same as the personalities in the training
data
10. 2 player minimax game
• Player 1 – Generator creates images
• Player 2 - Discriminator – recognizes the input as real or fake
• Adversarial competition on how to classify the fake samples generated by
generator . Generator tries to adapt the input to discriminator to cause it
to be misclassified. Discriminator tries to correctly classify fake as fake /
real as real
• NASH Equllibrium
• Generator recovers the data distribution correctly
• Discriminator random guess whether the input is real or fake
• Practically reaching NASH equilibrium is not possible but we have reached
to a place where we can generate realistic samples
11.
12. Generating face is relatively easy – BIG GAN
made it possible with imagenet
13. Advantage – able to learn with less
supervised
• Converting day scene to night scene
• Unsupervised image to image translation
20. QUANTUM GAN
• DENSITY MATRIX – STATES (possible
configuration of Q system
• Quantum distribution
• Quantum super imposition vs classical
uncertainity (undo is possible)