Deep Probabilistic Methods with PyTorch
summary
Starts from Bayesian linear regression and implements VI end to end in PyTorch 0.4, with notebooks on GitHub and Adam used for optimization. Derivations include Laplace, the ELBO objective, analytic KL terms and learning a full covariance via Cholesky. Covers VAEs (with the reparameterization trick), shows how discriminators encode density ratios, and reproduces a GAN mode‑collapse demo (generator collapses to one mode from an 8‑mode diamond). Expect concrete code, loss forms, and diagnostic plots rather than high‑level promises.
AI-written summary · about
chapters (11)
- 0:00Approximate Bayesian inference
- 10:43Posterior distributions in practice
- 17:28Conjugate models limitations
- 26:01Variational inference ELBO
- 33:08Variational methods neural nets
- 38:29Variational autoencoders
- 51:12Discriminators as generative models
- 58:26How GANs work
- 1:07:34Mode dropping in GANs
- 1:15:08Unifying VAE and GANs
- 1:19:50Probabilistic methods applications
related talks
01h 07mProbabilistic Programming and Readable Models
12mBayesian Neural Networks With Variational Inference in PyTorch
57mTutorial on Deep Learning III
01h 22mWriting effective bayesian programs using TensorFlow and TFP
37mVariational Inference in Python
35mVariational Inference and Python
21mProbabilistic Programming with Programmable Variational Inference
39mBayesian Deep Learning with Edward (and a trick using Dropout)