generativedeep-learningprobabilistic
Variational autoencoders
Probabilistic generative models optimising the ELBO — encode, sample in latent space, decode.
Уровни глубины
L0Intro~2ч
Knows VAE = encoder + decoder with stochastic bottleneck.
L1Basics~12ч
Derives ELBO and reparametrisation trick; trains VAE on MNIST.
L2Working~18ч
Builds β-VAE, conditional VAE; understands posterior collapse.
L3Advanced~25ч
Normalising flows, IWAE, discrete VAE (VQ-VAE); information-bottleneck connections.
L4Research~50ч
Hierarchical VAEs, disentanglement theory, score-based VAEs.
Ресурсы
L1 — Basics