\documentclass{article}
\usepackage{etoolbox}
\usepackage{fullpage}
\usepackage{color}
\usepackage{amsmath}
\usepackage{url}
\usepackage{verbatim}
\usepackage{graphicx}
\usepackage{parskip}
\usepackage{amssymb}
\usepackage{listings} % For displaying code
\begin{document}
\definecolor{blu}{rgb}{0,0,1}
\def\blu#1{{\color{blu}#1}}
\definecolor{gre}{rgb}{0,.5,0}
\def\gre#1{{\color{gre}#1}}
\definecolor{red}{rgb}{1,0,0}
\def\red#1{{\color{red}#1}}
\def\norm#1{\|#1\|}
\newcommand{\argmin}[1]{\mathop{\hbox{argmin}}_{#1}}
\newcommand{\argmax}[1]{\mathop{\hbox{argmax}}_{#1}}
\def\R{\mathbb{R}}
\newcommand{\fig}[2]{\includegraphics[width=#1\textwidth]{a5f/#2}}
\newcommand{\centerfig}[2]{\begin{center}\includegraphics[width=#1\textwidth]{a5f/#2}\end{center}}
\def\items#1{\begin{itemize}#1\end{itemize}}
\def\enum#1{\begin{enumerate}#1\end{enumerate}}
\def\argmax{\mathop{\rm arg\,max}}
\def\argmin{\mathop{\rm arg\,min}}
\def\half{\frac 1 2}
\newcommand{\code}[1]{\lstinputlisting[language=Matlab]{a5f/#1}}
\newcommand{\alignStar}[1]{\begin{align*}#1\end{align*}}
\newcommand{\mat}[1]{\begin{bmatrix}#1\end{bmatrix}}
\title{CPSC 540 Assignment 5 (due April 10)}
\author{UGMs, Bayes, and Literature Survey}
\date{}
\maketitle
\section{Undirected Graphical Models}
\subsection{Conditional UGM}
Consdier modeling the dependencies between sets of binary variables $x_j$ and $y_j$ with the following UGM which is a variation on a stacked RBM:
\centerfig{.6}{weirdRBM}
Computing univariate marginals in this model will be NP-hard in general, but the graph structure allows efficient block updates by conditioning on suitable subsets of the variables (this could be useful for designing approximate inference methods).
For each of the conditioning scenarios below, \blu{draw the conditional UGM and comment on how expensive it would be to compute univariate marginals} in the conditional UGM.
\enum{
\item Conditioning on all the $z$ and $h$ values.
\item Conditioning on all the $x$ and $h$ values.
\item Conditioning on all the $z$ and $y$ values.
\item Conditioning on all the $x$ and $z$ values.
}
\subsection{Fitting a UGM to PINs}
The function \emph{example\_UGM} loads a dataset $X$ containing samples of PIN numbers, based on the probabilities from the article at this URL: \url{http://www.datagenetics.com/blog/september32012}.\footnote{I got the probabilities from the reverse-engineered heatmap here: \url{http://jemore.free.fr/wordpress/?p=73}.}
This function shows how to use the UGM software to fit a UGM model to the dataset, where all node/edge parameters are tied and the graph is empty. It then performs decoding/inference/sampling in the fitted model. Unfortunately, this is not a very good model of the data for several reasons:
\enum{
\item The decoding is $1 \; 1 \; 1 \; 1$, whereas in the data the most likely value by far is $1 \; 2 \; 3\; 4$. Similarly, the sampler doesn't tend to generate $1 \; 2 \; 3\; 4$ even though this happens in more than 1/10 samples.
\item The marginal probability of the first number being $1$ is 22.06\%, which is acutally too low (it should be 38.54\%). In addition, the marginal probabilities of the remaining nubmers being $1$ are also 22.06\%, and these numbers are too high.
\item Conditioned on the first three numbers being $1 \; 2 \; 3$, the probability that the last number is 4 is less than 10\% in the model, whereas in the data it's more than 90\%.
}
In this question you'll explore better models of this data and different aspects of UGMs.
\enum{
\item \blu{What does $w$ have a length of $9$?}
\item \blu{Write an equation for the model being used by the code.}
\item \blu{What are potential sources of the problems above?}
\item Modify the demo to use a \emph{tied} value of 0 and re-run the demo. \blu{Why does the model now have $36$ parameters? Comment on whether this fixes each of the above 3 issues.}
\item Modify the demo to use chain-structured dependency (keeping the \emph{tied} value at 0). \blu{Comment on whether this fixes each of the above 3 issues.}
\item Modify the demo to use a completely-connected graph (keeping the \emph{tied} value at 0). \blu{Comment on whether this fixes each of the above 3 issues.}
\item UGM only support pairwise graphs, \blu{what would the effect of higher-order potentials be? What would the disdavantages of higher-order potentials be?}
}
If you want to further explore UGMs, there are quite a few demos on the UGM webpage that you can go through which cover all sorts of things like approximate inference and CRFs.
\section{Bayesian Inference}
\subsection{Conjugate Priors}
Consider a $y \in \{1,2,3\}$ following a multinoulli distribution with parameters $\theta = \{\theta_1,\theta_2,\theta_3\}$,
\[
y | \theta \sim \text{Mult}(\theta_1,\theta_2,\theta_3).
\]
We'll assume that $\theta$ follows a Dirichlet distribution (the conjugate prior to the multinoulli) with parameters $\alpha = \{\alpha_1,\alpha_2,\alpha_3\}$,
\[
\theta \sim \mathcal{D}(\alpha_1,\alpha_2,\alpha_3).
\]
Thus we have
\[
p(y|\theta,\alpha) = p(y|\theta) = \theta_1^{I(y=1)}\theta_2^{I(y=2)}\theta_3^{I(y=3)}, \quad p(\theta|\alpha) = \frac{\Gamma(\alpha_1+\alpha_2+\alpha_3)}{\Gamma(\alpha_1)\Gamma(\alpha_2)\Gamma(\alpha_3)}\theta_1^{\alpha_1-1}\theta_2^{\alpha_2-1}\theta_3^{\alpha_3-1}.
\]
Compute the following quantites:
\enum{
\item \blu{The posterior distribution,
\[
p(\theta|y,\alpha).
\]
\item \blu{The marginal likelihood of $y$ given the hyper-parameters $\alpha$,
\[
p(y|\alpha) = \int p(y,\theta|\alpha)d\theta,
\]}
\item \blu{The posterior mean estimate for $\theta$,
\[
\mathbb{E}_{\theta|y,\alpha}[\theta_i] = \int \theta_i p(\theta|y,\alpha)d\theta,
\]}
which (after some manipulation) should not involve any $\Gamma$ functions.
}
\item \blu{The posterior predictive distribution for a new independent observation $\hat{y}$ given $y$,
\[
p(\hat{y}|y,\alpha) = \int p(\hat{y},\theta|y,\alpha)d\theta.
\]
}
}
Hint: You can use $D(\alpha) = \frac{\Gamma(\alpha_1)\Gamma(\alpha_2)\Gamma(\alpha_3)}{\Gamma(\alpha_1+\alpha_2+\alpha_3)}$ to represent the normalizing constant of the prior and $D(\alpha^+)$ to give the normalizing constant of the posterior. You will also need to use that $\Gamma(\alpha+1) = \alpha\Gamma(\alpha)$. For some calculations you may find it a bit cleaner to parameterize the posterior in terms of $\beta_j = I(y=j) + \alpha_j$, and convert back once you have the final result.
\subsection{Empirical Bayes}
Consider the model
\[
y_i \sim \mathcal{N}(w^T\phi(x_i),\sigma^2), \quad w_j \sim \mathcal{N}(0,\lambda).
\]
By using properties of Gaussians the marginal likelihood has the form
\[
p(y_i|x_i,\sigma,\lambda) = (2\pi)^{-d/2}|C|^{-1/2}\exp\left(-\frac{y^TC^{-1}y}{2}\right),
\]
which gives a negative log-marginal likelihood of
\[
- \log p(y_i|x_i,\sigma,\lambda) \propto \log|C| + y^TC^{-1}y + \text{const.}
\]
where
\[
C = \frac{1}{\sigma^2}I + \frac{1}{\lambda}\Phi(X)\Phi(X)^T,
\]
As discussed in class, the marginal likelihood can be used to optimize hyper-parameters like $\sigma$, $\lambda$, and even the basis $\phi$.
The demo \emph{example\_basis} loads a dataset and fits a degree-2 polynomial to it. Normally we would use a test set to choose the degree of the polynomial but here we'll use the marginal likelihood of the training set. Write a function, \emph{leastSquaresEmpiricalBaysis}, that uses the marginal likelihood to choose the degree of the polynomial as well as the parameters $\lambda$ and $\sigma$ (you can assume that all $\lambda_j$ are equal, and you can restrict your search for $\lambda$ and $\sigma$ to powers of 10). \blu{Hand in your code and report the marginally most likely values of the degree, $\sigma$, and $\lambda$.} You can use the \emph{logdet} function to compute the log-determinant.
{Hint: computing $C^{-1}y$ by explicitly forming $C^{-1}$ may give you numerical issues that lead to non-sensical solution. You can avoid these by using $y^TC^{-1}y = y^Tv$ where $v$ is a solution to $Cv = y$ (Matlab will still give a warning due to ill-conditioning, but it won't return non-sensical results).}
\end{document}