Reading Assignment 1: Graphical Models & Monte Carlo Methods

This reading assignment covers required fundamentals that are essential for understanding the probabilistic approaches to deep learning we will focus on in this course.

From the Deep Learning Book - Chapter 16: Structured Probabilistic Models for Deep Learning, please read these sections:

• 16.1 The Challenge of Unstructured Modeling
  • Motivation for probabilistic (generative) models difference to supervised learning capabilities and use cases specific to generative models
• 16.2 Using Graphs to Describe Model Structure (16.2.1-5)
  (This is the most important section.)
  • Why do we need a partition function and how does that function relate to directed and undirected models?
  • What is the main advantage of energy-based models?
  • How is (free) energy related to the model’s capability to model data?
  • What does separation tell us about conditional dependencies?

• 16.3 Sampling from Graphical Models
  (This is the second most important section.)

• 16.4 Advantages of Structured Modeling

• 16.5 Learning about Dependencies

Sections 16.6 and 16.7 are optional. These topics will be covered later in depth.

Furthermore, please read Deep Learning Book - Chapter 17: Monte Carlo Methods with special emphasis on these sections:

• 17.1 Sampling and Monte Carlo Methods (for introduction)
  • Why do we need to sample?
  • What is the key idea about Monte Carlo?
  • 17.2: Importance sampling is optional
• 17.3 Markov Chain Monte Carlo Methods
  • What is MCMC and under which conditions it can be applied?
  • What is ‘burning in’ of markov chains?
• 17.4 Gibbs Sampling
  • How does Gibbs sampling work?
• 17.5 The Challenge of Mixing between Separated Modes
  • What is mixing and how is this related to MCMC methods? (no details)