Dive into the intricacies of Latent Variable Energy Based Models (LV-EBMs) and their training in this comprehensive lecture by Alfredo Canziani. Explore key concepts such as Zero Temperature Limit (ZTL) free energy, warmer free energy, and Infinite Temperature Limit. Examine practical examples of free energy calculations and gain insights into nomenclature and PyTorch implementation. Learn about loss functionals, compare ZTL and warmer temperature training approaches, and understand the conditional case. Visualize untrained and trained model manifolds, study energy functions, and discover the process of learning the Decoder. Conclude with valuable insights on choosing the appropriate latent size for your models.
Overview
Syllabus
– Welcome to class
– Zero Temperature Limit ZTL free energy recap
– warmer Free energy
– Infinite Temperature Limit
– Free energy, example y = Y[23]
– Free energy, example y = Y[10]
– Free energy, example y = 0, 0
– Nomenclature and PyTorch
– Training!
– Loss functionals
– ZTL vs. varmer temperature training
– Conditional case
– Untrained model manifold
– Energy function Ex, y, z
– Trained model manifold
– Learning the Decoder
– Choosing the latent size
– And that was it
Taught by
Alfredo Canziani
Tags
Related Courses
-
Latent Variable Energy Based Models - Inference
-
Joint Embedding Method and Latent Variable Energy Based Models
-
Energy Based Models and Self-Supervised Learning
-
Latent Variable EBMs for Structured Prediction
-
From LV-EBM to Target Prop to Autoencoder
-
Energy-Based Approaches to Representation Learning - Yann LeCun
