372 lines
14 KiB
TeX
372 lines
14 KiB
TeX
\documentclass[12pt,a4paper]{article}
|
|
|
|
\usepackage[utf8]{inputenc}
|
|
\usepackage[T1]{fontenc}
|
|
\usepackage{amsmath, amsfonts, amssymb, amsthm}
|
|
\usepackage{tikz}
|
|
\usepackage{listings}
|
|
\usepackage{fullpage}
|
|
|
|
|
|
\lstdefinelanguage{PseudoCode} {
|
|
morekeywords={
|
|
for,
|
|
while,
|
|
repeat,
|
|
from,
|
|
each,
|
|
foreach,
|
|
break,
|
|
continue,
|
|
in,
|
|
do,
|
|
as,
|
|
and,
|
|
or,
|
|
end,
|
|
return,
|
|
if,
|
|
then,
|
|
else,
|
|
function,
|
|
begin,
|
|
to,
|
|
new,
|
|
input,
|
|
output
|
|
},
|
|
morecomment=[l]{/*},
|
|
morecomment=[l]{//},
|
|
% basicstyle=\ttfamily,
|
|
% keywordstyle=\color{blue}, %\ttfamily,
|
|
commentstyle=\color{gray}\it,
|
|
keywordstyle=\bf,
|
|
rulecolor=\color{black},
|
|
literate=%
|
|
{!=}{{$\neq$}}1
|
|
{<=}{{$\leq$}}1
|
|
{>=}{{$\geq$}}1
|
|
{->}{{$\rightarrow$}}1
|
|
{<-}{{$\leftarrow$}}1
|
|
}
|
|
|
|
% },
|
|
% tabsize=3,
|
|
% sensitive=false,
|
|
% morecomment=[l]{#},
|
|
% morestring=[b]",
|
|
% extendedchars=true,
|
|
% inputencoding=utf8,
|
|
% literate=%
|
|
% {!=}{{$\neq$}}1
|
|
% {<=}{{$\leq$}}1
|
|
% {>=}{{$\geq$}}1
|
|
% {<>}{{$\neq$}}1
|
|
% {:=}{{$\ \leftarrow\quad$}}1
|
|
% {Ö}{{\"O}}1
|
|
% {Ä}{{\"A}}1
|
|
% {Ü}{{\"U}}1
|
|
% {ß}{{\ss{}}}1
|
|
% {ü}{{\"u}}1
|
|
% {ä}{{\"a}}1
|
|
% {ö}{{\"o}}1
|
|
% {~}{{\textasciitilde}}1,
|
|
% texcl=true % use all chars from \usepackage[utf8]{inputenc}
|
|
% }
|
|
\lstset{
|
|
tabsize=4,
|
|
xleftmargin=0pt, % left margin
|
|
numbers=left, % linenumber position
|
|
numbersep=15pt, % left linenumber padding
|
|
numberstyle=\tiny,
|
|
basicstyle=\ttfamily,
|
|
keywordstyle=\color{black!60},
|
|
commentstyle=\ttfamily\color{gray!70},
|
|
breaklines=true,
|
|
literate=
|
|
}
|
|
|
|
\renewcommand{\epsilon}{\varepsilon}
|
|
|
|
\newcommand{\vecl}{\ensuremath{\operatorname{vec}_l}}
|
|
\newcommand{\Sym}{\ensuremath{\operatorname{Sym}}}
|
|
|
|
\renewcommand{\vec}{\operatorname{vec}}
|
|
\newcommand{\devec}{\operatorname*{devec}}
|
|
\newcommand{\svec}{\operatorname{svec}}
|
|
\newcommand{\sym}{\operatorname{sym}}
|
|
\renewcommand{\skew}{\operatorname{skew}}
|
|
\newcommand{\rowSums}{\operatorname{rowSums}}
|
|
\newcommand{\colSums}{\operatorname{colSums}}
|
|
\newcommand{\diag}{\operatorname{diag}}
|
|
|
|
\begin{document}
|
|
|
|
\section{Kronecker Product Properties}
|
|
The \emph{mixed-product} property for matrices $A, B, C, D$ holds if and only if the following matrix products are well defined
|
|
\begin{displaymath}
|
|
(A\otimes B)(C \otimes D) = (A C) \otimes (B C).
|
|
\end{displaymath}
|
|
In combination with the \emph{Hadamard product} (element-wise multiplication) for matrices $A, C$ of the same size as well as $B, D$ of the same size is
|
|
\begin{displaymath}
|
|
(A\otimes B)\circ (C \otimes D) = (A \circ C) \otimes (B \circ D).
|
|
\end{displaymath}
|
|
The \emph{transpose} of the Kronecker product fulfills
|
|
\begin{displaymath}
|
|
(A\otimes B)^T = A^T \otimes B^T
|
|
\end{displaymath}
|
|
|
|
\section{Distance Computation}
|
|
The pair-wise distances $d_V(X_{i,:}, X_{j,:})$ arranged in the distance matrix $D\in\mathbb{R}^{n\times n}$ can be written as
|
|
\begin{align*}
|
|
\vec(D) = \rowSums(((X Q)\otimes 1_n - 1_n \otimes (X Q))^2)
|
|
\end{align*}
|
|
This can be computed in $\mathcal{O}(n^2p + np^2)$ time (vectorization and devectorization takes $\mathcal{O}(1)$).
|
|
|
|
The matrices $K, W$ are define through there elements as
|
|
\begin{displaymath}
|
|
k_{i j} = \exp\left(-\frac{d_{i j}^2}{2 h^2}\right),\qquad w_{i j} = \frac{k_{i j}}{\sum_{m} k_{m j}}.
|
|
\end{displaymath}
|
|
|
|
Next are $\bar{y}^{(m)}$ and the ``element-wise'' loss $l_i = L_n(V, X_i)$.
|
|
\begin{displaymath}
|
|
\bar{y}^{(m)} = W^T Y^m,\qquad l = \bar{y}^{(2)} - (\bar{y}^{(1)})^2
|
|
\end{displaymath}
|
|
|
|
\section{Gradient Computation}
|
|
The model
|
|
\begin{displaymath}
|
|
Y \sim g(B^T X) + \epsilon.
|
|
\end{displaymath}
|
|
|
|
Assume a data set $(X_i, Y_i)$ for $i = 1, ..., n$ with $X$ a $n\times p$ matrix such that each row represents one sample. Now let $l_i = L_n(V, X_i)$, $\bar{y}^{(1)}_j = (W^T Y)_j$ as well as $d_{i j}, w_{i j}$ the distance and weight matrix components. Then the gradient for the ``simple'' CVE method is given as
|
|
\begin{displaymath}
|
|
\nabla L_n(V) = \frac{1}{nh^2}\sum_{i = 1}^{n} \sum_{j = 1}^{n} (l_j - (Y_i - \bar{y}^{(1)}_j)^2) w_{i j} d_{i j} \nabla_V d_V(X_{i,:}, X_{j,:}).
|
|
\end{displaymath}
|
|
This representation is cumbersome and a direct implementation has a asymptotic run-time of $\Theta(n^2p^2)$ because it is a double sum over $n$, therefore quadratic in $n$, and the form of $\nabla_V d_V$.
|
|
|
|
This can be optimized and written in matrix notation. First the distance gradient is given as
|
|
\begin{displaymath}
|
|
\nabla_V d_V(X_{i,:}, X_{j,:}) = -2 (X_{i,:} - X_{j,:})^T (X_{i,:} - X_{j,:}) V
|
|
\end{displaymath}
|
|
(Note: $X_{i,:}\in\mathbb{R}^{1\times p}$, aka a row representing one sample). In addition define the $n\times n$ matrix $S$ through its elements
|
|
\begin{displaymath}
|
|
s_{i j} = (l_j - (Y_i - \bar{y}^{(1)}_j)^2) w_{i j} d_{i j}.
|
|
\end{displaymath}
|
|
Substitution in the gradient leads to
|
|
\begin{align*}
|
|
\nabla L_n(V)
|
|
&= -\frac{2}{nh^2}\sum_{i = 1}^{n} \sum_{j = 1}^{n} s_{i j} (X_{i,:} - X_{j,:})^T (X_{i,:} - X_{j,:}) V \\
|
|
&= -\frac{2}{nh^2}\sum_{i = 1}^{n} \sum_{j = 1}^{n} s_{i j} \left( X_{i,:}^T X_{i,:} - X_{i,:}^T X_{j,:} - X_{j,:}^T X_{i,:} + X_{j,:}^T X_{j,:} \right) V \\
|
|
&= -\frac{2}{nh^2} \left( \sum_{i = 1}^{n}\sum_{j = 1}^{n} (s_{i j} + s_{j i}) X_{i,:}^T X_{i,:} - \sum_{i = 1}^{n}\sum_{j = 1}^{n} (s_{i j} + s_{j i}) X_{i,:}^T X_{j,:} \right) V \\
|
|
&= -\frac{2}{nh^2} \left( X^T \diag(\colSums(S + S^T)) X - X^T (S + S^T) X \right) V \\
|
|
&= -\frac{2}{nh^2} X^T \left( \diag(\colSums(S + S^T)) - (S + S^T) \right) X V
|
|
\end{align*}
|
|
|
|
\begin{center}{\bf
|
|
ATTENTION: The given R examples are to illustrate the inplementation in C which is 0-indexed!
|
|
}\end{center}
|
|
|
|
The \emph{vertorization} operation maps a matrix $A\in\mathbb{R}^{n\times m}$ into $\mathbb{R}^{nm}$ by stacking the columns of $A$;
|
|
\begin{displaymath}
|
|
\vec(A) = (a_{0,0}, a_{0,1}, a_{0,2},...,a_{0,n-1},a_{1,0},a_{1,1},...,a_{n-1,n-1})^T.
|
|
\end{displaymath}
|
|
The relation $\vec(A)_k = a_{i,j}$ holds for $k=nj+i$ such that $0\leq k < n^2$ and $0\leq i < n, 0 \leq j < m$. This operation is obviously a bijection. When going ``backwards'' the dimension of the original space is required, therefore let $\devec_n$ be the operation such that $\devec_n(\vec(A)) = A$ for $A\in\mathbb{R}^{n\times m}$.\footnote{Note that for $B\in\mathbb{R}^{p\times q}$ such that $pq = nm$ the $\devec_n(\vec(B))\in\mathbb{R}^{n\times m}$.}
|
|
|
|
For symmetric matrices the information stored in $a_{i,j} = a_{j,i}$ is twice stored in $A=A^T\in\mathbb{R}^{n\times n}$, to remove this redundency the \emph{symmetric vectorization} is defined which saves the main diagonal and the lower triangular part of the symmetric matrix according the scema
|
|
\begin{displaymath}
|
|
\svec(A) = (a_{0,0},2a_{1,0},2a_{2,n},...,2a_{n-1,0},a_{1,1},2a_{2,1},...,2a_{n-1,1},a_{2,2},...,a_{n-1,n-1})
|
|
\end{displaymath}
|
|
A it more formal
|
|
\begin{displaymath}
|
|
\svec(A)_{k} = (2-\delta_{i,j})a_{i,j} \quad\text{for}\quad k = n j + i - \frac{j(j + 1)}{2}, 0\leq j \leq i < n^2.
|
|
\end{displaymath}
|
|
|
|
\begin{lstlisting}[language=R]
|
|
n <- 3
|
|
k <- function(i, j, n) { (j * n) + i - (j * (j + 1) / 2) }
|
|
i <- function(n) { rep(1:n - 1, n) }
|
|
j <- function(n) { rep(1:n - 1, each = n) }
|
|
A <- matrix(k(i(n), j(n), n), n)
|
|
A[which(j(n) > i(n))] <- NA
|
|
A
|
|
# [,1] [,2] [,3]
|
|
# [1,] 0 NA NA
|
|
# [2,] 1 3 NA
|
|
# [3,] 2 4 5
|
|
vec <- function(A) { as.vector(A) }
|
|
svec <- function(A) {
|
|
n <- nrow(A)
|
|
((2 - (i(n) == j(n))) * A)[i(n) >= j(n)]
|
|
}
|
|
svec(matrix(1, n, n))
|
|
# [1] 1 2 2 1 2 1
|
|
devec <- function(vec, n) { matrix(vec, n) }
|
|
\end{lstlisting}
|
|
|
|
For a quadratic matrix $A\in\mathbb{R}^{n\times n}$ we define
|
|
\begin{displaymath}
|
|
\sym(A) := \frac{A + A^T}{2}, \qquad \skew(A) := \frac{A - A^T}{2}.
|
|
\end{displaymath}
|
|
|
|
% For a Matrix $A\in\mathbb{R}^{n\times n}$ the \emph{vectorization} operation is defined as a mapping from the matrices into a
|
|
|
|
% Indexing a given matrix $A = (a_{ij})_{i,j = 1, ..., n} \in \mathbb{R}^{n\times n}$ given as
|
|
% \begin{displaymath}
|
|
% A = \begin{pmatrix}
|
|
% a_{0,0} & a_{0,1} & a_{0,2} & \ldots & a_{0,n-1} \\
|
|
% a_{1,0} & a_{1,1} & a_{1,2} & \ldots & a_{1,n-1} \\
|
|
% a_{2,0} & a_{2,1} & a_{2,2} & \ldots & a_{2,n-1} \\
|
|
% \vdots & \vdots & \vdots & \ddots & \vdots \\
|
|
% a_{n-1,0} & a_{n-1,1} & a_{n-1,2} & \ldots & a_{n-1,n-1}
|
|
% \end{pmatrix}
|
|
% \end{displaymath}
|
|
|
|
% A symmetric matrix with zero main diagonal, meaning a matrix $S = S^T$ with $S_{i,i} = 0,\ \forall i = 1,..,n$ is given in the following form
|
|
% \begin{displaymath}
|
|
% S = \begin{pmatrix}
|
|
% 0 & s_{1,0} & s_{2,0} & \ldots & s_{n-1,0} \\
|
|
% s_{1,0} & 0 & s_{2,1} & \ldots & s_{n-1,1} \\
|
|
% s_{2,0} & s_{2,1} & 0 & \ldots & s_{n-1,2} \\
|
|
% \vdots & \vdots & \vdots & \ddots & \vdots \\
|
|
% s_{n-1,0} & s_{n-1,1} & s_{n-1,2} & \ldots & 0
|
|
% \end{pmatrix}
|
|
% \end{displaymath}
|
|
% Therefore its sufficient to store only the lower triangular part, for memory efficiency and some further algorithmic shortcuts (sometime they are more expensive) the symmetric matrix $S$ is stored in packed form, meaning in a vector of the length $\frac{n(n-1)}{2}$. We use (like for matrices) a column-major order of elements and define the $\vecl:\Sym(n)\to \mathbb{R}^{n(n-1) / 2}$ operator defined as
|
|
|
|
% \begin{displaymath}
|
|
% \vecl(S) = (s_{1,0}, s_{2,0},\cdots,s_{n-1,0},s_{2,1}\cdots,s_{n-1,n-2})^T
|
|
% \end{displaymath}
|
|
|
|
% The relation between the matrix indices $i,j$ and the $\vecl$ index $k$ is given by
|
|
|
|
% \begin{displaymath}
|
|
% (\vecl(S)_k = s_{i,j} \quad\Leftrightarrow\quad k = jn+i) : j \in \{0,...,n-2\} \land j < i < n.
|
|
% \end{displaymath}
|
|
|
|
% \begin{center}
|
|
% \begin{tikzpicture}[xscale=1,yscale=-1]
|
|
% % \foreach \i in {0,...,5} {
|
|
% % \node at ({mod(\i, 3)}, {int(\i / 3)}) {$\i$};
|
|
% % }
|
|
% \foreach \i in {1,...,4} {
|
|
% \foreach \j in {1,...,\i} {
|
|
% \node at (\j, \i) {$\i,\j$};
|
|
% }
|
|
% }
|
|
|
|
% \end{tikzpicture}
|
|
% \end{center}
|
|
|
|
\newpage
|
|
\section{Algorithm}
|
|
The basic algorithm reads as follows:
|
|
|
|
Mit
|
|
\begin{displaymath}
|
|
X_{diff} := X\otimes 1_n - 1_n\otimes X
|
|
\end{displaymath}
|
|
gilt
|
|
\begin{displaymath}
|
|
X_{diff}Q := (X\otimes 1_n - 1_n\otimes X)Q = XQ\otimes 1_n - 1_n\otimes XQ
|
|
\end{displaymath}
|
|
|
|
\newcommand{\rStiefel}{\operatorname{rStiefel}}
|
|
% \lstset{language=PseudoCode}
|
|
% \begin{lstlisting}[mathescape, caption=Erste Phase von \texttt{HDE} (siehe \cite{HDE}), label=code:HDE, captionpos=b]
|
|
% \begin{lstlisting}[mathescape]
|
|
% // Hallo Welt
|
|
% /* Hallo comment */
|
|
% $X_{diff} \leftarrow X\otimes 1_n - 1_n\otimes X$
|
|
|
|
% for attempt from 1 to attempts do
|
|
% if $\exists V_{init}$ then
|
|
% $V \leftarrow V_{init}$
|
|
% else
|
|
% $V \leftarrow \rStiefel(p, q)$
|
|
% end if
|
|
|
|
% /* Projection matrix into null space */
|
|
% $Q \leftarrow I_p - VV^T$
|
|
|
|
% /* Pair-wise distances (row sum of squared elements) */
|
|
% $D \leftarrow$ foreach $i,j=1,...,n$ as $D_{i,j}\leftarrow \|(X_{i,:}-X_{j,:})Q\|_2^2$
|
|
|
|
% /* Weights */
|
|
% $W \leftarrow$ foreach $i,j=1,...,n$ as $W_{i,j} \leftarrow \frac{k(D_{i,j})}{\sum_{i} k(D_{i,j})}$
|
|
|
|
% $\bar{y}_1 \leftarrow W^TY$
|
|
% $\bar{y}_2 \leftarrow W^T(Y\odot Y)$
|
|
|
|
% /* Element-wise losses */
|
|
% $L \leftarrow \bar{y}_2 - \bar{y}_1^2$
|
|
|
|
% for epoch from 1 to epochs do
|
|
|
|
% $G_t \leftarrow \gamma G_{t-1} + (1-\gamma) \nabla_c L(V)$
|
|
|
|
% end for
|
|
% end for
|
|
% \end{lstlisting}
|
|
|
|
The loss at a given position is
|
|
\begin{displaymath}
|
|
L_n(V) = \frac{1}{nh^2}\sum_{i = 0}^{n - 1} \sum_{j = 0}^{n - 1} (L_j - (Y_i - \bar{y}^{(1)}_j)^2) w_{i j} d_{i j} \nabla_V d_V(X_{i,:}, X_{j,:})
|
|
\end{displaymath}
|
|
Now let the matrix $S$ be defined through its coefficients
|
|
\begin{displaymath}
|
|
s_{i j} = (L_j - (Y_i - \bar{y}^{(1)}_j)^2) w_{i j} d_{i j}
|
|
\end{displaymath}
|
|
This matrix is \underline{not} symmetric but we can consider the symmetric $S + S^T$ with a zero main diagonal because $D$ has a zero main diagonal, meaning $s_{i i} = 0$ because $d_{i i} = 0$ for each $i$. Therefore the following holds due to the fact that $\nabla_V d_V(X_{i,:}, X_{j,:}) = \nabla_V d_V(X_{j,:}, X_{i,:})$.
|
|
\begin{displaymath}
|
|
L_n(V) = \frac{1}{nh^2}\sum_{j = 0}^{n - 1} \sum_{i = j}^{n - 1} (s_{i j} + s_{j i}) \nabla_V d_V(X_{i,:}, X_{j,:})
|
|
\end{displaymath}
|
|
Note the summation indices $0 \leq j \leq i < n$. Substitution with $\nabla_V d_V(X_{i,:}, X_{j,:}) = -2 (X_{i,:} - X_{j,:})^T(X_{i,:} - X_{j,:}) V$ evaluates to
|
|
\begin{displaymath}
|
|
L_n(V) = -\frac{2}{nh^2}\sum_{j = 0}^{n - 1} \sum_{i = j}^{n - 1} (s_{i j} + s_{j i}) (X_{i,:} - X_{j,:})^T(X_{i,:} - X_{j,:}) V
|
|
\end{displaymath}
|
|
Let $X_{-}$ be the matrix containing all pairs of $X_{i,:}$ to $X_{j,:}$ differences using the same row indexing scheme as the symmetric vectorization.
|
|
\begin{displaymath}
|
|
(X_{-})_{k,:} = X_{i,:} - X_{j,:} \quad\text{for}\quad k = n j + i - \frac{j(j + 1)}{2}, 0\leq j \leq i < n^2
|
|
\end{displaymath}
|
|
With the $X_{-}$ matrix the above double sum can be formalized in matrix notation as follows\footnote{only valid cause $s_{i i} = 0$}
|
|
\begin{displaymath}
|
|
L_n(V) = -\frac{2}{nh^2} X_{-}^T(\svec(\sym(S)) \circ_r X_{-}) V
|
|
\end{displaymath}
|
|
where $\circ_r$ means the ``recycled'' hadamard product, this is for a vector $x\in\mathbb{R}^n$ and a Matrix $M\in\mathbb{R}^{n\times m}$ just the element wise multiplication for each column of $M$ with $x$, or equivalent $x\circ_r M = \underbrace{(x, x, ..., x)}_{{n\times m}} \circ M$ where $\circ$ is the element-wise product.
|
|
|
|
|
|
\begin{lstlisting}[mathescape, language=PseudoCode]
|
|
/* Starting value and initial gradient. */
|
|
$V_1 \leftarrow V_{init}$ if $\exists V_{init}$ else $\rStiefel(p, q)$
|
|
$G_1 \leftarrow (1 - \mu) \nabla L_n(V_0)$
|
|
|
|
/* Optimization loop */
|
|
$t \leftarrow 1$
|
|
while $t\leq\,$max.iter do
|
|
|
|
/* Update on stiefel manifold. */
|
|
$A \leftarrow G_tV_t^T - V_tG_t^T$
|
|
$V_{t+1} \leftarrow (I_p + \tau A)^{-1}(I_p - \tau A)V_{t}$
|
|
|
|
/* Check break condition. */
|
|
if $\|V_{t+1}V_{t+1}^T - V_{t}^TV_{t}\|_2^2 \leq \sqrt{2q}\,$tol then
|
|
break
|
|
end if
|
|
|
|
/* Check for decrease. */
|
|
if $L_n(V_{t+1}) - L_n(V_{t}) > L_n(V_{t})\,$slack then // TODO: slack?
|
|
/* Reduce step-size. */
|
|
$\tau \leftarrow \gamma\tau$
|
|
else
|
|
/* Gradient at next position (with momentum). */
|
|
$G_{t+1} \leftarrow \mu G_{t} + (1 - \mu) \nabla L_n(V_{t+1})$
|
|
/* Increase step index */
|
|
$t \leftarrow t + 1$
|
|
end if
|
|
|
|
end while
|
|
\end{lstlisting}
|
|
|
|
|
|
\end{document} |