Energy based models I
This lecture is a foundationational lecture for the concept of energy based models with a particular focus on the joint embedding method and latent variable energy based models 8LV-EBMs) and is a part of the Deep Learning Course at CDS, a course that covered the latest techniques in deep learning and representation learning, focusing on supervised and unsupervised deep learning, embedding methods, metric learning, convolutional and recurrent nets, with applications to computer vision, natural language understanding, and speech recognition. Prerequisites for this course include: Introduction to Deep Learning, Parameter sharing, and Introduction to Data Science or a Graduate Level Machine Learning.
Chapters:
00:00:00 – Welcome to class
00:00:39 – Predictive models
00:02:25 – Multi-output system
00:06:36 – Notation (factor graph)
00:07:41 – The energy function F(x, y)
00:08:53 – Inference
00:11:59 – Implicit function
00:15:53 – Conditional EBM
00:16:24 – Unconditional EBM
00:19:18 – EBM vs. probabilistic models
00:21:33 – Do we need a y at inference?
00:23:29 – When inference is hard
00:25:02 – Joint embeddings
00:28:29 – Latent variables
00:33:54 – Inference with latent variables
00:37:58 – Energies E and F
00:42:35 – Preview on the EBM practicum
00:44:30 – From energy to probabilities
00:50:37 – Examples: K-means and sparse coding
00:53:56 – Limiting the information capacity of the latent variable
00:57:24 – Training EBMs
01:04:02 – Maximum likelihood
01:13:58 – How to pick β?
01:17:28 – Problems with maximum likelihood
01:20:20 – Other types of loss functions
01:26:32 – Generalised margin loss
01:27:22 – General group loss
01:28:26 – Contrastive joint embeddings
01:34:51 – Denoising or mask autoencoder
01:46:14 – Summary and final remarks