ManySpace
Türkçe Sign in Sign up

AI & ML

Generative Models

Part of Generative Models, level 4. About 14 hours of study across the topics below.

Framing

  1. What Generative Models LearnModelling p(x): explicit versus implicit densities, sampling versus scoring, and the families that follow.1–2 h
  2. Evaluating Generative ModelsLikelihood, sample-based metrics (FID, precision/recall of distributions), and human judgment, and where each misleads.1–2 h

Latent Variable

  1. AutoencodersCompress then reconstruct: a learned bottleneck, and why a plain autoencoder is not yet a generative model.1–2 h
  2. Variational AutoencodersAn autoencoder trained as a latent-variable model: encoder as approximate posterior, loss as ELBO, reparameterisation for gradients.1–2 h
  3. Discrete Latents and Learned TokenizersVector quantisation turns continuous signals into codebook tokens that autoregressive and multimodal models can use.1–2 h

Adversarial

  1. Generative Adversarial NetworksA generator and a discriminator in a minimax game, and the instabilities (mode collapse, oscillation) that follow.1–2 h

Likelihood Based

  1. Normalizing FlowsInvertible transforms of a simple density with exact likelihood via the change-of-variables formula.1–2 h
  2. Energy-Based ModelsDensities defined by an unnormalised energy function, trained and sampled with Langevin dynamics or contrastive methods.1–2 h
  3. Autoregressive Generative ModelsGenerate one element at a time from a chain-rule factorisation: pixels, audio samples, or tokens.1–2 h

Diffusion

  1. Denoising Diffusion ModelsDestroy data with gradual Gaussian noise, learn to reverse it one step at a time.1–2 h
  2. Score-Based Models and the SDE ViewDiffusion as learning the score, the gradient of log density, and sampling by reversing a stochastic differential equation.1–2 h
  3. Flow MatchingTraining a velocity field that transports noise to data along simple paths, then sampling with an ODE solver.1–2 h
  4. Conditioning and GuidanceSteering generation with class labels or text: classifier guidance and classifier-free guidance.1–2 h
  5. Latent DiffusionRunning diffusion in a learned compressed latent space rather than pixel space.1–2 h