Learning Energy-Based Models of High-Dimensional Data - 2002

Explore a comprehensive lecture on energy-based models for high-dimensional data presented by Geoffry Hinton from the University of Toronto in 2002. Delve into the challenges of using belief networks for perception and discover an alternative approach using deterministic hidden units. Learn about the global energy concept and its role in defining data vector probabilities. Understand the limitations of maximum likelihood learning and uncover a new objective function for efficient parameter adjustment. Examine the application of these models to natural images, including filters and complex cells. Gain insights into unsupervised learning, various density models, Markov chains, and Hybrid Monte Carlo methods. This in-depth talk, hosted by the Center for Language & Speech Processing at Johns Hopkins University, offers a technical exploration of deterministic energy-based models and their potential in advancing machine learning and perception research.

Intro
Why unsupervised learning on data
Types of density models
Energybased model
Maximum likelihood learning
Markov chains
Hybrid Monte Carlo
The Obvious Algorithm
A View from Above
The Hard Side
Appliance to Images
Natural Images
Filters
Complex Cells
Technical Report
deterministic energybased models