D4Generative modeling

Generative modeling asks a model not to classify data but to produce it: to learn the underlying distribution well enough to sample new images, audio, or molecules that look like they belong. This is the machinery behind text-to-image systems and a deep probabilistic subject in its own right, where the central tension is always between sample quality, likelihood, and how stably the thing trains.

You will build autoencoders and their denoising variant, derive the VAE's evidence lower bound and reparameterization trick, work through GANs and their mode-collapse pitfalls, and use normalizing flows for exact likelihoods. The track then reaches diffusion: the forward noising process, the reverse denoising objective behind DDPM, latent diffusion as used in Stable Diffusion, and the newer flow-matching and consistency models that distill sampling down to a few steps.

These ideas drive image and video generation, drug discovery, and data augmentation, and they connect back to the Bayesian deep learning that shares the VAE's variational core.