Explore the intricacies of Latent Variable Energy Based Models (LV-EBMs) with a focus on inference in this comprehensive lecture. Delve into topics such as affine transformations, training sample mappings, and the unconditional case. Discover the energy function and its applications, including indexing by individual training samples and analyzing specific energy shapes. Examine the concept of free energy, its definition, and practical examples. Learn how to compute free energy across the entire space. Gain valuable insights into LV-EBMs through clear explanations and visual demonstrations using a graphic tablet.
Latent Variable Energy Based Models - Inference
Overview
Syllabus
– Affine transformation in 2 and 3D by @LeiosLabs James Schloss
– Thanks for sending me a Wacom graphic tablet
– *Inference* for LV EBM we're given a model
– Training samples: one to many mapping
– Let's simplify stuff: the unconditional case
– Untrained model manifold generation
– The Energy Function, tadaaa
– Indexing energy function by picking individual training samples
– The 23rd energy U shaped
– The 10th energy ~ shaped
– The Free Energy definition and the 10th example
– The 23rd free energy
– Computing the free energy for the entire space
– That was it :
Taught by
Alfredo Canziani
