Explore joint embedding methods and latent variable energy-based models (LV-EBMs) in this comprehensive lecture by Yann LeCun. Delve into predictive models, multi-output systems, and factor graph notation before examining energy functions and inference processes. Investigate conditional and unconditional EBMs, comparing them to probabilistic models. Learn about joint embeddings, latent variables, and their role in inference. Discover how to convert energy to probabilities and explore examples like K-means and sparse coding. Understand the training process for EBMs, including maximum likelihood and alternative loss functions. Examine contrastive joint embeddings and denoising autoencoders. Gain valuable insights into advanced machine learning concepts and techniques throughout this in-depth presentation.
Joint Embedding Method and Latent Variable Energy Based Models
Overview
Syllabus
– Welcome to class
– Predictive models
– Multi-output system
– Notation factor graph
– The energy function Fx, y
– Inference
– Implicit function
– Conditional EBM
– Unconditional EBM
– EBM vs. probabilistic models
– Do we need a y at inference?
– When inference is hard
– Joint embeddings
– Latent variables
– Inference with latent variables
– Energies E and F
– Preview on the EBM practicum
– From energy to probabilities
– Examples: K-means and sparse coding
– Limiting the information capacity of the latent variable
– Training EBMs
– Maximum likelihood
– How to pick β?
– Problems with maximum likelihood
– Other types of loss functions
– Generalised margin loss
– General group loss
– Contrastive joint embeddings
– Denoising or mask autoencoder
– Summary and final remarks
Taught by
Alfredo Canziani
