Week 1 · Day 5 · Friday 16 October

Generative AI for Vision: GANs

Generative adversarial networks: the generator-discriminator game and its objective, training and its failure modes, conditional GANs, and unpaired image-to-image translation with CycleGAN.

Topic Description & Objectives Resource
Generative Adversarial Networks The generator and the discriminator, the minimax objective, training, GAN variants, latent space interpolation, and the problems with GANs Open
Lab 1: GANs on Anime Faces Build a generator and a discriminator and train them against each other to generate anime faces Notebook Colab
Lab 2: Conditional GANs on CelebA Control what the generator draws by conditioning both networks on face attributes Notebook Colab
Lab 3: CycleGAN, Monet to Photo Unpaired translation between Monet paintings and photographs, with a PatchGAN discriminator, cycle-consistency loss and a replay buffer Notebook Colab
Optional: Conditional GANs on Animal Faces Generate cats, dogs or wild animals on demand by conditioning on the class label Notebook Colab