Latent Variable Models

William James Hall, Harvard University

Monday 01/12/2015, WH1305, 10am-5pm

Karthik Dinakar

Massachusetts Institute of Technology

k...@media.mit.edu


The purpose of this workshop is to bring clinical psychology and probabilistic graphical models closer to each other. Such a deep interdisciplinary collaboration can generate an array of insights that are symmetric to both fields. First, you will be introduced to machine learning - we will look at classes of models and how they are built and evaluated. Next, we will focus on a family of models called latent variable models and study how they are practically built. We will concretize the aforementioned concepts and see how they are being applied to model self-injury. Finally, we will take an honest look at inhibitors and catalysts for this deep collaboration to reach fruition.

Introduction

10:00-12:00pm

Model

  • What is a model?
  • How are models useful?
  • Real world examples
  • Jargon alert!

Types of Models I

  • Supervised learning
  • Unsupervised learning
  • Structured learning
  • Active learning

Types of Models II

  • Discriminative
  • Generative
  • Frequentist
  • Bayesian ✓

Techniques

12:45-2:30pm

Box's Loop ++

  • Model preparation
  • Model parameterization
  • Model selection
  • Model criticism
  • Model augmentation
  • Humans in the loop ✓✓

Key tools

  • Contextual Inquiry & Design
  • Sampling
  • Etic/Emic Qualitatitive Analysis
  • Cognitive Walkthroughs
  • Heuristic evaluations
  • Decision-theoretic models
  • Data Visualization

Modeling self-harm

2:50-4:00pm

We will go through a step-by-step walkthrough of using latent variable models for modeling, understanding and potentially predicting self-harm. We will deliberate on the use of the above tools, explore ways of dealing with sparsity of variables and ask how hierarchical Poisson matrix factorization and model criticism in particular can help us understand self-harm. Readers should ask about De Finette's theorem of exchangablity.

Cultural inhibitors & catalysts

4:15-5:00pm

Inhibitors

  • Physics-pretense
  • Physics-envy
  • The Post-Facto Critic
  • Structural impediments in academia

Catalysts

  • Human-in-the-loop approx. posterior inference
  • Model augmentation
  • Celebration of canonical cases

References

Frank & Witten
Data Mining: Practical Machine Learning Tools and Techniques Book
Christopher M.Bishop
Pattern Recognition and Machine Learning Book
Kevin P. Murphy
Machine Learning: A Probabilistic Perspective (MIT Press) Book
Daphne Koller and Nir Friedman
Probabilistic Graphical Models: Principles and Techniques (MIT Press) Book
David M Blei
Writes beautifully and lucidly, and is a leader in the field of latent variable models Annual Review Article

Generated from a haughty shell script 01/06/2015 13:26