📄️ Generative AI
A premium course on generative modelling: build a VAE and a DCGAN on MNIST, a diffusion model on CelebA, evaluate with FID and think through provenance and responsible use.
📄️ 1. Discriminative vs generative
Module 1 of the Generative AI premium course: modelling p(y|x) versus p(x), likelihood, sampling, and a chronology of the three main families.
📄️ 2. Autoencoders
Module 2 of the Generative AI premium course: the bottleneck, reconstruction, interpolation in latent space, and why a plain autoencoder generates badly.
📄️ 3. Variational autoencoders
Module 3 of the Generative AI premium course: latent distribution, reparameterisation trick, KL term, why reconstructions are blurry, and a VAE trained on MNIST.
📄️ 4. GAN training
Module 4 of the Generative AI premium course: adversarial losses, alternating training, DCGAN on MNIST and reading loss curves that mean less than they seem.
📄️ 5. Mode collapse
Module 5 of the Generative AI premium course: symptoms of mode collapse, Wasserstein loss and gradient penalty, spectral normalisation, and a StyleGAN glimpse.
📄️ 6. Diffusion models
Module 6 of the Generative AI premium course: forward and reverse processes, noise prediction, U-Net conditioned on the timestep, DDPM and DDIM sampling, latent diffusion.
📄️ 7. Text conditioning
Module 7 of the Generative AI premium course: text encoders like CLIP, cross-attention, classifier-free guidance, the guidance scale, and the Stable Diffusion link.
📄️ 8. FID and metrics
Module 8 of the Generative AI premium course: FID and inception score statistics, their limits, precision-recall for generation and the role of human evaluation.
📄️ 9. Text, image, audio
Module 9 of the Generative AI premium course: autoregressive versus diffusion depending on the modality, available models, generation cost and a multimodality preview.
📄️ 10. Copyright and responsible use
Module 10 of the Generative AI premium course: training data and licences, watermarks and C2PA, deepfakes, consent, and the regulation front-facing systems must apply.
📄️ Recap and exam
Complete recap of the Generative AI premium course: VAE, GAN, diffusion, text conditioning, metrics, modalities and responsible use, then the 40-question exam.