PPAML Summer School 2016

Each day is about 3hrs of teaching. (The first day maybe a little longer?)

Day 1

###Part 1: Introduction

• Resources
  • This is a mashup of the introductory chapters of dippl, probmods, and agentmodels.
    • http://dippl.org/chapters/02-webppl.html
    • https://probmods.org/generative-models.html
    • https://probmods.org/conditioning.html
    • https://probmods.org/patterns-of-inference.html
    • http://agentmodels.org/chapters/02-webppl.html
• Algorithms
  • Rejection
  • Enumeration
• Content
  • Getting set up
    • Browser
    • Node
  • Generative models
  • The WebPPL language
    • ERPs as reified distributions
    • Visualizing distributions
    • Examples that illustrate the language
      • Coin flipping
      • Geometric distribution as example with structure change
        • Recursion
  • Conditioning and factors
  • Inference operators
  • Graphical models as programs with fixed control flow
  • More interesting (but still simple) models
    • Tug of war
    • A simple data science example? Bayesian regression?

###Part 2: A tour through model-space

• Resources
  • This is based partly on dippl, and partly on Forest, although we'll need to translate many of the models to webppl.
• Algorithms
  • Particle filters
  • Basic MCMC
• Content
  • Markov models
    • -> HMMs
      • -> PCFGs
  • Hierarchical models
    • Bags of marbles
  • Mixture models
    • Topic models
  • Logistic regression
  • Bayesian neural net
  • ...others?

##Day 3: Bayesian data analysis and making predictions from data (3.5h)

• Resources
  • This is a shortened version of MH's BDA course.
  • http://forestdb.org/models/bayesian-data-analysis.html
    • In contrast to this page, we might want 2-3 real datasets
• Algorithms
  • MCMC
    • Incremental
    • HMC
  • Variational inference
• Content
  • MH to revise content
  • Input/output of data
  • Occam's razor
  • Various models useful for BDA
  • Use BDA to compare two of the models shown earlier on a dataset
    • logistic regression vs Bayesian neural net
    • a rich cognitive model vs regression
    • text analysis models (topic models, hmms, etc)

Day 4: Agents (2.5h)

• Resources
  • This is based on agentmodels.org, which by then will hopefully be fairly smooth.
• Algorithms
  • Enumeration + caching
  • Nested inference
• Content
  • Reasoning about reasoning via nested conditioning
  • Multi-agent reasoning
  • Language understanding
  • Modeling agents
    • Planning as inference
    • MDPs
      • • visualizations
    • POMDPS
    • Inferring beliefs, preferences, and other properties from observed behavior